@guidekit/vad 0.1.0-beta.1 → 0.1.0-beta.2

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 GuideKit
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/dist/index.cjs

@@ -38,7 +38,7 @@ __export(index_exports, {
 });
 module.exports = __toCommonJS(index_exports);
 var ort = __toESM(require("onnxruntime-web"), 1);
-var VAD_VERSION = "0.1.0";
+var VAD_VERSION = "0.1.0-beta.2";
 var LOG_PREFIX = "[GuideKit:VAD]";
 var DEFAULT_MODEL_URL = "https://cdn.jsdelivr.net/npm/@ricky0123/vad-web@0.0.20/dist/silero_vad_v5.onnx";
 var CACHE_NAME = `guidekit-vad-v${VAD_VERSION}`;
@@ -230,9 +230,9 @@ var SileroVAD = class {
     this._resetStates();
     this._isCalibrating = true;
     this._calibrationSamples = [];
-    this._calibrationFramesNeeded = Math.ceil(
+    this._calibrationFramesNeeded = Math.max(1, Math.floor(
       CALIBRATION_DURATION_MS / 1e3 * this._sampleRate / FRAME_SIZE
-    );
+    ));
     this._log("Calibrating noise floor for", this._calibrationFramesNeeded, "frames");
     this._setupAudioPipeline(stream);
   }

package/dist/index.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/index.ts"],"sourcesContent":["// @guidekit/vad — Silero VAD ONNX model wrapper for voice activity detection\nimport * as ort from 'onnxruntime-web';\n\nexport const VAD_VERSION = '0.1.0';\n\n// ---------------------------------------------------------------------------\n// Constants\n// ---------------------------------------------------------------------------\n\nconst LOG_PREFIX = '[GuideKit:VAD]';\n\n/** Default CDN URL for the Silero VAD ONNX model (v5). */\nconst DEFAULT_MODEL_URL =\n  'https://cdn.jsdelivr.net/npm/@ricky0123/vad-web@0.0.20/dist/silero_vad_v5.onnx';\n\n/** Cache API key used for persisting the downloaded ONNX model. */\nconst CACHE_NAME = `guidekit-vad-v${VAD_VERSION}`;\nconst CACHE_MODEL_KEY = 'model.onnx';\n\n/** Silero VAD frame size: 512 samples at 16 kHz = 32 ms per frame. */\nconst FRAME_SIZE = 512;\n\n/** Target sample rate for VAD processing. */\nconst TARGET_SAMPLE_RATE = 16000;\n\n/** Duration (in ms) of audio collected for noise floor calibration. */\nconst CALIBRATION_DURATION_MS = 500;\n\n/** Hidden/cell state size for Silero VAD v5 LSTM. */\nconst STATE_SIZE = 128;\n\n// ---------------------------------------------------------------------------\n// Types\n// ---------------------------------------------------------------------------\n\nexport interface VADOptions {\n  /** Speech probability threshold (0-1). Default: 0.5 */\n  threshold?: number;\n  /** Minimum speech duration in ms to trigger start. Default: 300 */\n  minSpeechDurationMs?: number;\n  /** Silence duration in ms after speech to trigger end. Default: 500 */\n  silenceDurationMs?: number;\n  /** Sample rate. Default: 16000 */\n  sampleRate?: number;\n  /** Enable debug logging. Default: false */\n  debug?: boolean;\n  /** Custom URL for the Silero VAD ONNX model file. */\n  modelUrl?: string;\n}\n\nexport interface VADEvent {\n  type: 'speech-start' | 'speech-end' | 'vad-ready';\n  timestamp: number;\n  /** Speech probability (0-1) at the moment of the event. */\n  probability?: number;\n}\n\ntype VADEventType = VADEvent['type'];\ntype VADCallback = (event: VADEvent) => void;\n\n// ---------------------------------------------------------------------------\n// Utility: Cache API helpers\n// ---------------------------------------------------------------------------\n\nasync function loadModelFromCache(): Promise<ArrayBuffer | null> {\n  if (typeof caches === 'undefined') return null;\n  try {\n    const cache = await caches.open(CACHE_NAME);\n    const response = await cache.match(CACHE_MODEL_KEY);\n    return response ? response.arrayBuffer() : null;\n  } catch {\n    // Cache API may be unavailable in certain contexts (e.g. opaque origins).\n    return null;\n  }\n}\n\nasync function saveModelToCache(data: ArrayBuffer): Promise<void> {\n  if (typeof caches === 'undefined') return;\n  try {\n    const cache = await caches.open(CACHE_NAME);\n    await cache.put(CACHE_MODEL_KEY, new Response(data));\n  } catch {\n    // Silently ignore cache write failures.\n  }\n}\n\n// ---------------------------------------------------------------------------\n// Utility: Resampler\n// ---------------------------------------------------------------------------\n\n/**\n * Simple linear-interpolation resampler from `inputRate` to `outputRate`.\n * Adequate for VAD where perceptual audio quality is irrelevant.\n */\nfunction resample(\n  input: Float32Array,\n  inputRate: number,\n  outputRate: number,\n): Float32Array {\n  if (inputRate === outputRate) return input;\n  const ratio = inputRate / outputRate;\n  const outputLength = Math.round(input.length / ratio);\n  const output = new Float32Array(outputLength);\n  for (let i = 0; i < outputLength; i++) {\n    const srcIndex = i * ratio;\n    const srcFloor = Math.floor(srcIndex);\n    const srcCeil = Math.min(srcFloor + 1, input.length - 1);\n    const frac = srcIndex - srcFloor;\n    output[i] = (input[srcFloor] as number) * (1 - frac) + (input[srcCeil] as number) * frac;\n  }\n  return output;\n}\n\n// ---------------------------------------------------------------------------\n// SileroVAD\n// ---------------------------------------------------------------------------\n\nexport class SileroVAD {\n  // Options (resolved with defaults)\n  private readonly _threshold: number;\n  private readonly _minSpeechDurationMs: number;\n  private readonly _silenceDurationMs: number;\n  private readonly _sampleRate: number;\n  private readonly _debug: boolean;\n  private readonly _modelUrl: string;\n\n  // ONNX Runtime state\n  private _session: ort.InferenceSession | null = null;\n  private _h: ort.Tensor | null = null;\n  private _c: ort.Tensor | null = null;\n\n  // Audio pipeline\n  private _audioContext: AudioContext | null = null;\n  private _ownsAudioContext = false;\n  private _sourceNode: MediaStreamAudioSourceNode | null = null;\n  private _workletNode: AudioWorkletNode | ScriptProcessorNode | null = null;\n  private _stream: MediaStream | null = null;\n\n  // Frame buffer for accumulating resampled samples into FRAME_SIZE chunks\n  private _frameBuffer: Float32Array = new Float32Array(0);\n  private _frameBufferOffset = 0;\n\n  // State tracking\n  private _isReady = false;\n  private _isSpeaking = false;\n  private _isStarted = false;\n  private _isDestroyed = false;\n\n  // Duration tracking (in frames)\n  private _consecutiveSpeechFrames = 0;\n  private _consecutiveSilenceFrames = 0;\n  private _frameDurationMs: number;\n  private _minSpeechFrames: number;\n  private _silenceFrames: number;\n\n  // Noise floor calibration\n  private _isCalibrating = false;\n  private _calibrationSamples: number[] = [];\n  private _calibrationFramesNeeded = 0;\n  private _calibratedThreshold: number;\n\n  // Event listeners\n  private _listeners: Map<VADEventType, Set<VADCallback>> = new Map();\n\n  // Processing lock to serialise frame inference\n  private _processingPromise: Promise<void> = Promise.resolve();\n\n  constructor(options?: VADOptions) {\n    this._threshold = options?.threshold ?? 0.5;\n    this._minSpeechDurationMs = options?.minSpeechDurationMs ?? 300;\n    this._silenceDurationMs = options?.silenceDurationMs ?? 500;\n    this._sampleRate = options?.sampleRate ?? TARGET_SAMPLE_RATE;\n    this._debug = options?.debug ?? false;\n    this._modelUrl = options?.modelUrl ?? DEFAULT_MODEL_URL;\n    this._calibratedThreshold = this._threshold;\n\n    // Pre-compute frame-duration-based counters\n    this._frameDurationMs = (FRAME_SIZE / this._sampleRate) * 1000;\n    this._minSpeechFrames = Math.ceil(this._minSpeechDurationMs / this._frameDurationMs);\n    this._silenceFrames = Math.ceil(this._silenceDurationMs / this._frameDurationMs);\n\n    this._log('Created with options', {\n      threshold: this._threshold,\n      minSpeechDurationMs: this._minSpeechDurationMs,\n      silenceDurationMs: this._silenceDurationMs,\n      sampleRate: this._sampleRate,\n      modelUrl: this._modelUrl,\n    });\n  }\n\n  // -------------------------------------------------------------------------\n  // Public API\n  // -------------------------------------------------------------------------\n\n  /** Load the ONNX model. Uses Cache API for persistence across sessions. */\n  async init(): Promise<void> {\n    if (this._isDestroyed) {\n      throw new Error(`${LOG_PREFIX} Cannot init after destroy`);\n    }\n    if (this._isReady) {\n      this._log('Already initialised — skipping');\n      return;\n    }\n\n    this._log('Initialising...');\n\n    // 1. Attempt to load model bytes from cache, falling back to network.\n    let modelBuffer = await loadModelFromCache();\n    if (modelBuffer) {\n      this._log('Loaded model from Cache API');\n    } else {\n      this._log('Fetching model from', this._modelUrl);\n      const response = await fetch(this._modelUrl);\n      if (!response.ok) {\n        throw new Error(\n          `${LOG_PREFIX} Failed to fetch model: ${response.status} ${response.statusText}`,\n        );\n      }\n      modelBuffer = await response.arrayBuffer();\n      this._log('Model fetched, size:', modelBuffer.byteLength, 'bytes');\n\n      // Persist to Cache API for next time.\n      await saveModelToCache(modelBuffer);\n      this._log('Model saved to Cache API');\n    }\n\n    // 2. Create ONNX InferenceSession.\n    this._session = await ort.InferenceSession.create(modelBuffer, {\n      executionProviders: ['wasm'],\n      graphOptimizationLevel: 'all',\n    });\n\n    // 3. Initialise LSTM hidden/cell state tensors (zeros).\n    this._resetStates();\n\n    this._isReady = true;\n    this._log('Model loaded and ready');\n\n    this._emit({\n      type: 'vad-ready',\n      timestamp: Date.now(),\n    });\n  }\n\n  /**\n   * Process a single audio frame (512 samples at 16 kHz).\n   * Returns the speech probability (0-1).\n   */\n  async processFrame(audioData: Float32Array): Promise<number> {\n    if (!this._session) {\n      throw new Error(`${LOG_PREFIX} Model not loaded. Call init() first.`);\n    }\n    if (audioData.length !== FRAME_SIZE) {\n      throw new Error(\n        `${LOG_PREFIX} Expected ${FRAME_SIZE} samples, got ${audioData.length}`,\n      );\n    }\n\n    const inputTensor = new ort.Tensor('float32', audioData, [1, FRAME_SIZE]);\n    const srTensor = new ort.Tensor('int64', BigInt64Array.from([BigInt(this._sampleRate)]), [1]);\n\n    const feeds: Record<string, ort.Tensor> = {\n      input: inputTensor,\n      sr: srTensor,\n      h: this._h!,\n      c: this._c!,\n    };\n\n    const results = await this._session.run(feeds);\n\n    // Update LSTM hidden/cell states for the next frame.\n    this._h = results['hn'] as ort.Tensor;\n    this._c = results['cn'] as ort.Tensor;\n\n    const probability = (results['output'] as ort.Tensor).data[0] as number;\n    return probability;\n  }\n\n  /** Start VAD processing on a MediaStream (typically from getUserMedia). */\n  start(stream: MediaStream): void {\n    if (this._isDestroyed) {\n      throw new Error(`${LOG_PREFIX} Cannot start after destroy`);\n    }\n    if (!this._isReady) {\n      throw new Error(`${LOG_PREFIX} Model not loaded. Call init() first.`);\n    }\n    if (this._isStarted) {\n      this._log('Already started — stopping previous session first');\n      this.stop();\n    }\n\n    this._log('Starting VAD on MediaStream');\n    this._stream = stream;\n    this._isStarted = true;\n\n    // Reset speech tracking state.\n    this._isSpeaking = false;\n    this._consecutiveSpeechFrames = 0;\n    this._consecutiveSilenceFrames = 0;\n    this._frameBuffer = new Float32Array(FRAME_SIZE);\n    this._frameBufferOffset = 0;\n\n    // Reset LSTM states for a fresh stream.\n    this._resetStates();\n\n    // Begin noise floor calibration.\n    this._isCalibrating = true;\n    this._calibrationSamples = [];\n    this._calibrationFramesNeeded = Math.ceil(\n      (CALIBRATION_DURATION_MS / 1000) * this._sampleRate / FRAME_SIZE,\n    );\n    this._log('Calibrating noise floor for', this._calibrationFramesNeeded, 'frames');\n\n    // Build the audio processing pipeline.\n    this._setupAudioPipeline(stream);\n  }\n\n  /** Stop VAD processing and release audio resources (but keep the model). */\n  stop(): void {\n    if (!this._isStarted) return;\n\n    this._log('Stopping VAD');\n\n    // Tear down audio nodes.\n    this._teardownAudioPipeline();\n\n    // If we were speaking, emit speech-end.\n    if (this._isSpeaking) {\n      this._isSpeaking = false;\n      this._emit({\n        type: 'speech-end',\n        timestamp: Date.now(),\n        probability: 0,\n      });\n    }\n\n    // Reset state.\n    this._isStarted = false;\n    this._isSpeaking = false;\n    this._consecutiveSpeechFrames = 0;\n    this._consecutiveSilenceFrames = 0;\n    this._frameBufferOffset = 0;\n    this._isCalibrating = false;\n    this._calibrationSamples = [];\n    this._processingPromise = Promise.resolve();\n\n    this._log('VAD stopped');\n  }\n\n  /** Register a callback for speech-start events. Returns an unsubscribe function. */\n  onSpeechStart(callback: VADCallback): () => void {\n    return this._on('speech-start', callback);\n  }\n\n  /** Register a callback for speech-end events. Returns an unsubscribe function. */\n  onSpeechEnd(callback: VADCallback): () => void {\n    return this._on('speech-end', callback);\n  }\n\n  /** Register a callback for vad-ready events. Returns an unsubscribe function. */\n  onReady(callback: VADCallback): () => void {\n    return this._on('vad-ready', callback);\n  }\n\n  /** Whether the ONNX model is loaded and ready. */\n  get isReady(): boolean {\n    return this._isReady;\n  }\n\n  /** Whether speech is currently detected. */\n  get isSpeaking(): boolean {\n    return this._isSpeaking;\n  }\n\n  /** The MediaStream currently being processed, or null. */\n  get stream(): MediaStream | null {\n    return this._stream;\n  }\n\n  /** Release ONNX model session and all audio resources. */\n  async destroy(): Promise<void> {\n    if (this._isDestroyed) return;\n    this._log('Destroying...');\n\n    this.stop();\n\n    if (this._session) {\n      await this._session.release();\n      this._session = null;\n    }\n\n    // Dispose tensors.\n    this._h?.dispose();\n    this._c?.dispose();\n    this._h = null;\n    this._c = null;\n\n    this._isReady = false;\n    this._isDestroyed = true;\n    this._listeners.clear();\n\n    this._log('Destroyed');\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: Event system\n  // -------------------------------------------------------------------------\n\n  private _on(type: VADEventType, callback: VADCallback): () => void {\n    let set = this._listeners.get(type);\n    if (!set) {\n      set = new Set();\n      this._listeners.set(type, set);\n    }\n    set.add(callback);\n    return () => {\n      set!.delete(callback);\n    };\n  }\n\n  private _emit(event: VADEvent): void {\n    const set = this._listeners.get(event.type);\n    if (!set) return;\n    for (const cb of set) {\n      try {\n        cb(event);\n      } catch (err) {\n        // eslint-disable-next-line no-console\n        console.error(`${LOG_PREFIX} Error in ${event.type} callback:`, err);\n      }\n    }\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: Audio pipeline\n  // -------------------------------------------------------------------------\n\n  private _setupAudioPipeline(stream: MediaStream): void {\n    // Determine the incoming sample rate.\n    const tracks = stream.getAudioTracks();\n    const trackSettings = tracks[0]?.getSettings();\n    const inputSampleRate = trackSettings?.sampleRate ?? 48000;\n\n    this._log('Input sample rate:', inputSampleRate);\n\n    // Create AudioContext at the input sample rate so we don't double-resample.\n    // SSR guard: AudioContext may not exist.\n    if (typeof AudioContext === 'undefined' && typeof webkitAudioContext === 'undefined') {\n      throw new Error(`${LOG_PREFIX} AudioContext is not available in this environment`);\n    }\n\n    const AudioContextClass =\n      typeof AudioContext !== 'undefined'\n        ? AudioContext\n        : // eslint-disable-next-line @typescript-eslint/no-explicit-any\n          (globalThis as any).webkitAudioContext as typeof AudioContext;\n\n    this._audioContext = new AudioContextClass({ sampleRate: inputSampleRate });\n    this._ownsAudioContext = true;\n\n    this._sourceNode = this._audioContext.createMediaStreamSource(stream);\n\n    // Try AudioWorklet first, fall back to ScriptProcessorNode.\n    this._setupScriptProcessor(inputSampleRate);\n  }\n\n  /**\n   * ScriptProcessorNode fallback (works everywhere, including Safari).\n   * We use a buffer size of 4096 which gives ~85 ms of audio at 48 kHz.\n   */\n  private _setupScriptProcessor(inputSampleRate: number): void {\n    if (!this._audioContext || !this._sourceNode) return;\n\n    // Buffer size must be a power of 2: 256, 512, 1024, 2048, 4096, 8192, 16384.\n    const bufferSize = 4096;\n    const processor = this._audioContext.createScriptProcessor(bufferSize, 1, 1);\n\n    processor.onaudioprocess = (event: AudioProcessingEvent) => {\n      if (!this._isStarted) return;\n\n      const inputData = event.inputBuffer.getChannelData(0);\n\n      // Resample to target rate if needed.\n      const resampled =\n        inputSampleRate !== this._sampleRate\n          ? resample(inputData, inputSampleRate, this._sampleRate)\n          : new Float32Array(inputData);\n\n      // Feed resampled audio into frame-sized chunks.\n      this._feedAudio(resampled);\n    };\n\n    this._sourceNode.connect(processor);\n    processor.connect(this._audioContext.destination);\n    this._workletNode = processor;\n\n    this._log('Audio pipeline set up (ScriptProcessorNode)');\n  }\n\n  /**\n   * Accumulate resampled audio into FRAME_SIZE chunks and process each full frame.\n   */\n  private _feedAudio(samples: Float32Array): void {\n    let offset = 0;\n\n    while (offset < samples.length) {\n      const remaining = FRAME_SIZE - this._frameBufferOffset;\n      const available = samples.length - offset;\n      const toCopy = Math.min(remaining, available);\n\n      this._frameBuffer.set(\n        samples.subarray(offset, offset + toCopy),\n        this._frameBufferOffset,\n      );\n      this._frameBufferOffset += toCopy;\n      offset += toCopy;\n\n      if (this._frameBufferOffset === FRAME_SIZE) {\n        const frame = new Float32Array(this._frameBuffer);\n        this._frameBufferOffset = 0;\n\n        // Serialise inference calls to avoid overlapping ONNX sessions.\n        this._processingPromise = this._processingPromise.then(() =>\n          this._handleFrame(frame),\n        );\n      }\n    }\n  }\n\n  /**\n   * Process a single FRAME_SIZE frame: run inference and update speech state.\n   */\n  private async _handleFrame(frame: Float32Array): Promise<void> {\n    if (!this._isStarted || !this._session) return;\n\n    let probability: number;\n    try {\n      probability = await this.processFrame(frame);\n    } catch (err) {\n      if (this._debug) {\n        // eslint-disable-next-line no-console\n        console.error(`${LOG_PREFIX} Inference error:`, err);\n      }\n      return;\n    }\n\n    // Noise floor calibration phase.\n    if (this._isCalibrating) {\n      this._calibrationSamples.push(probability);\n\n      if (this._calibrationSamples.length >= this._calibrationFramesNeeded) {\n        this._finishCalibration();\n      }\n      return;\n    }\n\n    // Speech state machine.\n    const isSpeechFrame = probability >= this._calibratedThreshold;\n\n    if (isSpeechFrame) {\n      this._consecutiveSpeechFrames++;\n      this._consecutiveSilenceFrames = 0;\n\n      if (!this._isSpeaking && this._consecutiveSpeechFrames >= this._minSpeechFrames) {\n        this._isSpeaking = true;\n        this._log('Speech started, probability:', probability.toFixed(3));\n        this._emit({\n          type: 'speech-start',\n          timestamp: Date.now(),\n          probability,\n        });\n      }\n    } else {\n      this._consecutiveSilenceFrames++;\n      // Do NOT reset _consecutiveSpeechFrames here — only reset when speech-end fires.\n\n      if (this._isSpeaking && this._consecutiveSilenceFrames >= this._silenceFrames) {\n        this._isSpeaking = false;\n        this._consecutiveSpeechFrames = 0;\n        this._log('Speech ended, probability:', probability.toFixed(3));\n        this._emit({\n          type: 'speech-end',\n          timestamp: Date.now(),\n          probability,\n        });\n      }\n    }\n  }\n\n  private _finishCalibration(): void {\n    if (this._calibrationSamples.length === 0) {\n      this._isCalibrating = false;\n      return;\n    }\n\n    // Compute average noise floor probability.\n    const sum = this._calibrationSamples.reduce((a, b) => a + b, 0);\n    const avgNoise = sum / this._calibrationSamples.length;\n\n    // If the ambient noise floor is high, nudge the threshold above it.\n    // We add a margin so we don't constantly trigger on background noise.\n    const NOISE_MARGIN = 0.15;\n    if (avgNoise + NOISE_MARGIN > this._threshold) {\n      this._calibratedThreshold = Math.min(avgNoise + NOISE_MARGIN, 0.95);\n      this._log(\n        'Noise floor is high. Adjusted threshold from',\n        this._threshold.toFixed(3),\n        'to',\n        this._calibratedThreshold.toFixed(3),\n        '(avg noise:',\n        avgNoise.toFixed(3) + ')',\n      );\n    } else {\n      this._calibratedThreshold = this._threshold;\n      this._log('Noise floor OK, avg:', avgNoise.toFixed(3), '— keeping threshold at', this._threshold.toFixed(3));\n    }\n\n    // Recompute frame counters in case threshold changed min speech behaviour.\n    this._minSpeechFrames = Math.ceil(this._minSpeechDurationMs / this._frameDurationMs);\n    this._silenceFrames = Math.ceil(this._silenceDurationMs / this._frameDurationMs);\n\n    this._isCalibrating = false;\n    this._calibrationSamples = [];\n  }\n\n  private _teardownAudioPipeline(): void {\n    if (this._workletNode) {\n      try {\n        this._workletNode.disconnect();\n      } catch {\n        // Ignore disconnect errors.\n      }\n      if ('onaudioprocess' in this._workletNode) {\n        (this._workletNode as ScriptProcessorNode).onaudioprocess = null;\n      }\n      this._workletNode = null;\n    }\n\n    if (this._sourceNode) {\n      try {\n        this._sourceNode.disconnect();\n      } catch {\n        // Ignore.\n      }\n      this._sourceNode = null;\n    }\n\n    if (this._audioContext && this._ownsAudioContext) {\n      try {\n        void this._audioContext.close();\n      } catch {\n        // Ignore.\n      }\n      this._audioContext = null;\n      this._ownsAudioContext = false;\n    }\n\n    this._stream = null;\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: ONNX state helpers\n  // -------------------------------------------------------------------------\n\n  /** Reset the LSTM hidden and cell states to zeros. */\n  private _resetStates(): void {\n    // Dispose any existing tensors to free memory.\n    this._h?.dispose();\n    this._c?.dispose();\n\n    const zeros = new Float32Array(2 * STATE_SIZE).fill(0);\n    this._h = new ort.Tensor('float32', zeros.slice(0, STATE_SIZE), [2, 1, 64]);\n    this._c = new ort.Tensor('float32', zeros.slice(STATE_SIZE), [2, 1, 64]);\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: Logging\n  // -------------------------------------------------------------------------\n\n  private _log(...args: unknown[]): void {\n    if (!this._debug) return;\n    // eslint-disable-next-line no-console\n    console.log(LOG_PREFIX, ...args);\n  }\n}\n\n// ---------------------------------------------------------------------------\n// Convenience factory\n// ---------------------------------------------------------------------------\n\n/**\n * Create and initialise a SileroVAD instance in one call.\n *\n * ```ts\n * const vad = await createVAD({ debug: true });\n * vad.onSpeechStart(() => console.log('speaking'));\n * vad.start(stream);\n * ```\n */\nexport async function createVAD(options?: VADOptions): Promise<SileroVAD> {\n  const vad = new SileroVAD(options);\n  await vad.init();\n  return vad;\n}\n\n// Re-export the frame size constant so consumers can align their buffers.\nexport { FRAME_SIZE, TARGET_SAMPLE_RATE };\n\n// Type-only declaration for environments that provide webkitAudioContext.\ndeclare global {\n  // eslint-disable-next-line no-var\n  var webkitAudioContext: typeof AudioContext | undefined;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AACA,UAAqB;AAEd,IAAM,cAAc;AAM3B,IAAM,aAAa;AAGnB,IAAM,oBACJ;AAGF,IAAM,aAAa,iBAAiB,WAAW;AAC/C,IAAM,kBAAkB;AAGxB,IAAM,aAAa;AAGnB,IAAM,qBAAqB;AAG3B,IAAM,0BAA0B;AAGhC,IAAM,aAAa;AAmCnB,eAAe,qBAAkD;AAC/D,MAAI,OAAO,WAAW,YAAa,QAAO;AAC1C,MAAI;AACF,UAAM,QAAQ,MAAM,OAAO,KAAK,UAAU;AAC1C,UAAM,WAAW,MAAM,MAAM,MAAM,eAAe;AAClD,WAAO,WAAW,SAAS,YAAY,IAAI;AAAA,EAC7C,QAAQ;AAEN,WAAO;AAAA,EACT;AACF;AAEA,eAAe,iBAAiB,MAAkC;AAChE,MAAI,OAAO,WAAW,YAAa;AACnC,MAAI;AACF,UAAM,QAAQ,MAAM,OAAO,KAAK,UAAU;AAC1C,UAAM,MAAM,IAAI,iBAAiB,IAAI,SAAS,IAAI,CAAC;AAAA,EACrD,QAAQ;AAAA,EAER;AACF;AAUA,SAAS,SACP,OACA,WACA,YACc;AACd,MAAI,cAAc,WAAY,QAAO;AACrC,QAAM,QAAQ,YAAY;AAC1B,QAAM,eAAe,KAAK,MAAM,MAAM,SAAS,KAAK;AACpD,QAAM,SAAS,IAAI,aAAa,YAAY;AAC5C,WAAS,IAAI,GAAG,IAAI,cAAc,KAAK;AACrC,UAAM,WAAW,IAAI;AACrB,UAAM,WAAW,KAAK,MAAM,QAAQ;AACpC,UAAM,UAAU,KAAK,IAAI,WAAW,GAAG,MAAM,SAAS,CAAC;AACvD,UAAM,OAAO,WAAW;AACxB,WAAO,CAAC,IAAK,MAAM,QAAQ,KAAgB,IAAI,QAAS,MAAM,OAAO,IAAe;AAAA,EACtF;AACA,SAAO;AACT;AAMO,IAAM,YAAN,MAAgB;AAAA;AAAA,EAEJ;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA;AAAA,EAGT,WAAwC;AAAA,EACxC,KAAwB;AAAA,EACxB,KAAwB;AAAA;AAAA,EAGxB,gBAAqC;AAAA,EACrC,oBAAoB;AAAA,EACpB,cAAiD;AAAA,EACjD,eAA8D;AAAA,EAC9D,UAA8B;AAAA;AAAA,EAG9B,eAA6B,IAAI,aAAa,CAAC;AAAA,EAC/C,qBAAqB;AAAA;AAAA,EAGrB,WAAW;AAAA,EACX,cAAc;AAAA,EACd,aAAa;AAAA,EACb,eAAe;AAAA;AAAA,EAGf,2BAA2B;AAAA,EAC3B,4BAA4B;AAAA,EAC5B;AAAA,EACA;AAAA,EACA;AAAA;AAAA,EAGA,iBAAiB;AAAA,EACjB,sBAAgC,CAAC;AAAA,EACjC,2BAA2B;AAAA,EAC3B;AAAA;AAAA,EAGA,aAAkD,oBAAI,IAAI;AAAA;AAAA,EAG1D,qBAAoC,QAAQ,QAAQ;AAAA,EAE5D,YAAY,SAAsB;AAChC,SAAK,aAAa,SAAS,aAAa;AACxC,SAAK,uBAAuB,SAAS,uBAAuB;AAC5D,SAAK,qBAAqB,SAAS,qBAAqB;AACxD,SAAK,cAAc,SAAS,cAAc;AAC1C,SAAK,SAAS,SAAS,SAAS;AAChC,SAAK,YAAY,SAAS,YAAY;AACtC,SAAK,uBAAuB,KAAK;AAGjC,SAAK,mBAAoB,aAAa,KAAK,cAAe;AAC1D,SAAK,mBAAmB,KAAK,KAAK,KAAK,uBAAuB,KAAK,gBAAgB;AACnF,SAAK,iBAAiB,KAAK,KAAK,KAAK,qBAAqB,KAAK,gBAAgB;AAE/E,SAAK,KAAK,wBAAwB;AAAA,MAChC,WAAW,KAAK;AAAA,MAChB,qBAAqB,KAAK;AAAA,MAC1B,mBAAmB,KAAK;AAAA,MACxB,YAAY,KAAK;AAAA,MACjB,UAAU,KAAK;AAAA,IACjB,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,OAAsB;AAC1B,QAAI,KAAK,cAAc;AACrB,YAAM,IAAI,MAAM,GAAG,UAAU,4BAA4B;AAAA,IAC3D;AACA,QAAI,KAAK,UAAU;AACjB,WAAK,KAAK,qCAAgC;AAC1C;AAAA,IACF;AAEA,SAAK,KAAK,iBAAiB;AAG3B,QAAI,cAAc,MAAM,mBAAmB;AAC3C,QAAI,aAAa;AACf,WAAK,KAAK,6BAA6B;AAAA,IACzC,OAAO;AACL,WAAK,KAAK,uBAAuB,KAAK,SAAS;AAC/C,YAAM,WAAW,MAAM,MAAM,KAAK,SAAS;AAC3C,UAAI,CAAC,SAAS,IAAI;AAChB,cAAM,IAAI;AAAA,UACR,GAAG,UAAU,2BAA2B,SAAS,MAAM,IAAI,SAAS,UAAU;AAAA,QAChF;AAAA,MACF;AACA,oBAAc,MAAM,SAAS,YAAY;AACzC,WAAK,KAAK,wBAAwB,YAAY,YAAY,OAAO;AAGjE,YAAM,iBAAiB,WAAW;AAClC,WAAK,KAAK,0BAA0B;AAAA,IACtC;AAGA,SAAK,WAAW,MAAU,qBAAiB,OAAO,aAAa;AAAA,MAC7D,oBAAoB,CAAC,MAAM;AAAA,MAC3B,wBAAwB;AAAA,IAC1B,CAAC;AAGD,SAAK,aAAa;AAElB,SAAK,WAAW;AAChB,SAAK,KAAK,wBAAwB;AAElC,SAAK,MAAM;AAAA,MACT,MAAM;AAAA,MACN,WAAW,KAAK,IAAI;AAAA,IACtB,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,aAAa,WAA0C;AAC3D,QAAI,CAAC,KAAK,UAAU;AAClB,YAAM,IAAI,MAAM,GAAG,UAAU,uCAAuC;AAAA,IACtE;AACA,QAAI,UAAU,WAAW,YAAY;AACnC,YAAM,IAAI;AAAA,QACR,GAAG,UAAU,aAAa,UAAU,iBAAiB,UAAU,MAAM;AAAA,MACvE;AAAA,IACF;AAEA,UAAM,cAAc,IAAQ,WAAO,WAAW,WAAW,CAAC,GAAG,UAAU,CAAC;AACxE,UAAM,WAAW,IAAQ,WAAO,SAAS,cAAc,KAAK,CAAC,OAAO,KAAK,WAAW,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;AAE5F,UAAM,QAAoC;AAAA,MACxC,OAAO;AAAA,MACP,IAAI;AAAA,MACJ,GAAG,KAAK;AAAA,MACR,GAAG,KAAK;AAAA,IACV;AAEA,UAAM,UAAU,MAAM,KAAK,SAAS,IAAI,KAAK;AAG7C,SAAK,KAAK,QAAQ,IAAI;AACtB,SAAK,KAAK,QAAQ,IAAI;AAEtB,UAAM,cAAe,QAAQ,QAAQ,EAAiB,KAAK,CAAC;AAC5D,WAAO;AAAA,EACT;AAAA;AAAA,EAGA,MAAM,QAA2B;AAC/B,QAAI,KAAK,cAAc;AACrB,YAAM,IAAI,MAAM,GAAG,UAAU,6BAA6B;AAAA,IAC5D;AACA,QAAI,CAAC,KAAK,UAAU;AAClB,YAAM,IAAI,MAAM,GAAG,UAAU,uCAAuC;AAAA,IACtE;AACA,QAAI,KAAK,YAAY;AACnB,WAAK,KAAK,wDAAmD;AAC7D,WAAK,KAAK;AAAA,IACZ;AAEA,SAAK,KAAK,6BAA6B;AACvC,SAAK,UAAU;AACf,SAAK,aAAa;AAGlB,SAAK,cAAc;AACnB,SAAK,2BAA2B;AAChC,SAAK,4BAA4B;AACjC,SAAK,eAAe,IAAI,aAAa,UAAU;AAC/C,SAAK,qBAAqB;AAG1B,SAAK,aAAa;AAGlB,SAAK,iBAAiB;AACtB,SAAK,sBAAsB,CAAC;AAC5B,SAAK,2BAA2B,KAAK;AAAA,MAClC,0BAA0B,MAAQ,KAAK,cAAc;AAAA,IACxD;AACA,SAAK,KAAK,+BAA+B,KAAK,0BAA0B,QAAQ;AAGhF,SAAK,oBAAoB,MAAM;AAAA,EACjC;AAAA;AAAA,EAGA,OAAa;AACX,QAAI,CAAC,KAAK,WAAY;AAEtB,SAAK,KAAK,cAAc;AAGxB,SAAK,uBAAuB;AAG5B,QAAI,KAAK,aAAa;AACpB,WAAK,cAAc;AACnB,WAAK,MAAM;AAAA,QACT,MAAM;AAAA,QACN,WAAW,KAAK,IAAI;AAAA,QACpB,aAAa;AAAA,MACf,CAAC;AAAA,IACH;AAGA,SAAK,aAAa;AAClB,SAAK,cAAc;AACnB,SAAK,2BAA2B;AAChC,SAAK,4BAA4B;AACjC,SAAK,qBAAqB;AAC1B,SAAK,iBAAiB;AACtB,SAAK,sBAAsB,CAAC;AAC5B,SAAK,qBAAqB,QAAQ,QAAQ;AAE1C,SAAK,KAAK,aAAa;AAAA,EACzB;AAAA;AAAA,EAGA,cAAc,UAAmC;AAC/C,WAAO,KAAK,IAAI,gBAAgB,QAAQ;AAAA,EAC1C;AAAA;AAAA,EAGA,YAAY,UAAmC;AAC7C,WAAO,KAAK,IAAI,cAAc,QAAQ;AAAA,EACxC;AAAA;AAAA,EAGA,QAAQ,UAAmC;AACzC,WAAO,KAAK,IAAI,aAAa,QAAQ;AAAA,EACvC;AAAA;AAAA,EAGA,IAAI,UAAmB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,aAAsB;AACxB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,SAA6B;AAC/B,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,MAAM,UAAyB;AAC7B,QAAI,KAAK,aAAc;AACvB,SAAK,KAAK,eAAe;AAEzB,SAAK,KAAK;AAEV,QAAI,KAAK,UAAU;AACjB,YAAM,KAAK,SAAS,QAAQ;AAC5B,WAAK,WAAW;AAAA,IAClB;AAGA,SAAK,IAAI,QAAQ;AACjB,SAAK,IAAI,QAAQ;AACjB,SAAK,KAAK;AACV,SAAK,KAAK;AAEV,SAAK,WAAW;AAChB,SAAK,eAAe;AACpB,SAAK,WAAW,MAAM;AAEtB,SAAK,KAAK,WAAW;AAAA,EACvB;AAAA;AAAA;AAAA;AAAA,EAMQ,IAAI,MAAoB,UAAmC;AACjE,QAAI,MAAM,KAAK,WAAW,IAAI,IAAI;AAClC,QAAI,CAAC,KAAK;AACR,YAAM,oBAAI,IAAI;AACd,WAAK,WAAW,IAAI,MAAM,GAAG;AAAA,IAC/B;AACA,QAAI,IAAI,QAAQ;AAChB,WAAO,MAAM;AACX,UAAK,OAAO,QAAQ;AAAA,IACtB;AAAA,EACF;AAAA,EAEQ,MAAM,OAAuB;AACnC,UAAM,MAAM,KAAK,WAAW,IAAI,MAAM,IAAI;AAC1C,QAAI,CAAC,IAAK;AACV,eAAW,MAAM,KAAK;AACpB,UAAI;AACF,WAAG,KAAK;AAAA,MACV,SAAS,KAAK;AAEZ,gBAAQ,MAAM,GAAG,UAAU,aAAa,MAAM,IAAI,cAAc,GAAG;AAAA,MACrE;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAMQ,oBAAoB,QAA2B;AAErD,UAAM,SAAS,OAAO,eAAe;AACrC,UAAM,gBAAgB,OAAO,CAAC,GAAG,YAAY;AAC7C,UAAM,kBAAkB,eAAe,cAAc;AAErD,SAAK,KAAK,sBAAsB,eAAe;AAI/C,QAAI,OAAO,iBAAiB,eAAe,OAAO,uBAAuB,aAAa;AACpF,YAAM,IAAI,MAAM,GAAG,UAAU,oDAAoD;AAAA,IACnF;AAEA,UAAM,oBACJ,OAAO,iBAAiB,cACpB;AAAA;AAAA,MAEC,WAAmB;AAAA;AAE1B,SAAK,gBAAgB,IAAI,kBAAkB,EAAE,YAAY,gBAAgB,CAAC;AAC1E,SAAK,oBAAoB;AAEzB,SAAK,cAAc,KAAK,cAAc,wBAAwB,MAAM;AAGpE,SAAK,sBAAsB,eAAe;AAAA,EAC5C;AAAA;AAAA;AAAA;AAAA;AAAA,EAMQ,sBAAsB,iBAA+B;AAC3D,QAAI,CAAC,KAAK,iBAAiB,CAAC,KAAK,YAAa;AAG9C,UAAM,aAAa;AACnB,UAAM,YAAY,KAAK,cAAc,sBAAsB,YAAY,GAAG,CAAC;AAE3E,cAAU,iBAAiB,CAAC,UAAgC;AAC1D,UAAI,CAAC,KAAK,WAAY;AAEtB,YAAM,YAAY,MAAM,YAAY,eAAe,CAAC;AAGpD,YAAM,YACJ,oBAAoB,KAAK,cACrB,SAAS,WAAW,iBAAiB,KAAK,WAAW,IACrD,IAAI,aAAa,SAAS;AAGhC,WAAK,WAAW,SAAS;AAAA,IAC3B;AAEA,SAAK,YAAY,QAAQ,SAAS;AAClC,cAAU,QAAQ,KAAK,cAAc,WAAW;AAChD,SAAK,eAAe;AAEpB,SAAK,KAAK,6CAA6C;AAAA,EACzD;AAAA;AAAA;AAAA;AAAA,EAKQ,WAAW,SAA6B;AAC9C,QAAI,SAAS;AAEb,WAAO,SAAS,QAAQ,QAAQ;AAC9B,YAAM,YAAY,aAAa,KAAK;AACpC,YAAM,YAAY,QAAQ,SAAS;AACnC,YAAM,SAAS,KAAK,IAAI,WAAW,SAAS;AAE5C,WAAK,aAAa;AAAA,QAChB,QAAQ,SAAS,QAAQ,SAAS,MAAM;AAAA,QACxC,KAAK;AAAA,MACP;AACA,WAAK,sBAAsB;AAC3B,gBAAU;AAEV,UAAI,KAAK,uBAAuB,YAAY;AAC1C,cAAM,QAAQ,IAAI,aAAa,KAAK,YAAY;AAChD,aAAK,qBAAqB;AAG1B,aAAK,qBAAqB,KAAK,mBAAmB;AAAA,UAAK,MACrD,KAAK,aAAa,KAAK;AAAA,QACzB;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,MAAc,aAAa,OAAoC;AAC7D,QAAI,CAAC,KAAK,cAAc,CAAC,KAAK,SAAU;AAExC,QAAI;AACJ,QAAI;AACF,oBAAc,MAAM,KAAK,aAAa,KAAK;AAAA,IAC7C,SAAS,KAAK;AACZ,UAAI,KAAK,QAAQ;AAEf,gBAAQ,MAAM,GAAG,UAAU,qBAAqB,GAAG;AAAA,MACrD;AACA;AAAA,IACF;AAGA,QAAI,KAAK,gBAAgB;AACvB,WAAK,oBAAoB,KAAK,WAAW;AAEzC,UAAI,KAAK,oBAAoB,UAAU,KAAK,0BAA0B;AACpE,aAAK,mBAAmB;AAAA,MAC1B;AACA;AAAA,IACF;AAGA,UAAM,gBAAgB,eAAe,KAAK;AAE1C,QAAI,eAAe;AACjB,WAAK;AACL,WAAK,4BAA4B;AAEjC,UAAI,CAAC,KAAK,eAAe,KAAK,4BAA4B,KAAK,kBAAkB;AAC/E,aAAK,cAAc;AACnB,aAAK,KAAK,gCAAgC,YAAY,QAAQ,CAAC,CAAC;AAChE,aAAK,MAAM;AAAA,UACT,MAAM;AAAA,UACN,WAAW,KAAK,IAAI;AAAA,UACpB;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF,OAAO;AACL,WAAK;AAGL,UAAI,KAAK,eAAe,KAAK,6BAA6B,KAAK,gBAAgB;AAC7E,aAAK,cAAc;AACnB,aAAK,2BAA2B;AAChC,aAAK,KAAK,8BAA8B,YAAY,QAAQ,CAAC,CAAC;AAC9D,aAAK,MAAM;AAAA,UACT,MAAM;AAAA,UACN,WAAW,KAAK,IAAI;AAAA,UACpB;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,qBAA2B;AACjC,QAAI,KAAK,oBAAoB,WAAW,GAAG;AACzC,WAAK,iBAAiB;AACtB;AAAA,IACF;AAGA,UAAM,MAAM,KAAK,oBAAoB,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC9D,UAAM,WAAW,MAAM,KAAK,oBAAoB;AAIhD,UAAM,eAAe;AACrB,QAAI,WAAW,eAAe,KAAK,YAAY;AAC7C,WAAK,uBAAuB,KAAK,IAAI,WAAW,cAAc,IAAI;AAClE,WAAK;AAAA,QACH;AAAA,QACA,KAAK,WAAW,QAAQ,CAAC;AAAA,QACzB;AAAA,QACA,KAAK,qBAAqB,QAAQ,CAAC;AAAA,QACnC;AAAA,QACA,SAAS,QAAQ,CAAC,IAAI;AAAA,MACxB;AAAA,IACF,OAAO;AACL,WAAK,uBAAuB,KAAK;AACjC,WAAK,KAAK,wBAAwB,SAAS,QAAQ,CAAC,GAAG,+BAA0B,KAAK,WAAW,QAAQ,CAAC,CAAC;AAAA,IAC7G;AAGA,SAAK,mBAAmB,KAAK,KAAK,KAAK,uBAAuB,KAAK,gBAAgB;AACnF,SAAK,iBAAiB,KAAK,KAAK,KAAK,qBAAqB,KAAK,gBAAgB;AAE/E,SAAK,iBAAiB;AACtB,SAAK,sBAAsB,CAAC;AAAA,EAC9B;AAAA,EAEQ,yBAA+B;AACrC,QAAI,KAAK,cAAc;AACrB,UAAI;AACF,aAAK,aAAa,WAAW;AAAA,MAC/B,QAAQ;AAAA,MAER;AACA,UAAI,oBAAoB,KAAK,cAAc;AACzC,QAAC,KAAK,aAAqC,iBAAiB;AAAA,MAC9D;AACA,WAAK,eAAe;AAAA,IACtB;AAEA,QAAI,KAAK,aAAa;AACpB,UAAI;AACF,aAAK,YAAY,WAAW;AAAA,MAC9B,QAAQ;AAAA,MAER;AACA,WAAK,cAAc;AAAA,IACrB;AAEA,QAAI,KAAK,iBAAiB,KAAK,mBAAmB;AAChD,UAAI;AACF,aAAK,KAAK,cAAc,MAAM;AAAA,MAChC,QAAQ;AAAA,MAER;AACA,WAAK,gBAAgB;AACrB,WAAK,oBAAoB;AAAA,IAC3B;AAEA,SAAK,UAAU;AAAA,EACjB;AAAA;AAAA;AAAA;AAAA;AAAA,EAOQ,eAAqB;AAE3B,SAAK,IAAI,QAAQ;AACjB,SAAK,IAAI,QAAQ;AAEjB,UAAM,QAAQ,IAAI,aAAa,IAAI,UAAU,EAAE,KAAK,CAAC;AACrD,SAAK,KAAK,IAAQ,WAAO,WAAW,MAAM,MAAM,GAAG,UAAU,GAAG,CAAC,GAAG,GAAG,EAAE,CAAC;AAC1E,SAAK,KAAK,IAAQ,WAAO,WAAW,MAAM,MAAM,UAAU,GAAG,CAAC,GAAG,GAAG,EAAE,CAAC;AAAA,EACzE;AAAA;AAAA;AAAA;AAAA,EAMQ,QAAQ,MAAuB;AACrC,QAAI,CAAC,KAAK,OAAQ;AAElB,YAAQ,IAAI,YAAY,GAAG,IAAI;AAAA,EACjC;AACF;AAeA,eAAsB,UAAU,SAA0C;AACxE,QAAM,MAAM,IAAI,UAAU,OAAO;AACjC,QAAM,IAAI,KAAK;AACf,SAAO;AACT;","names":[]}
+{"version":3,"sources":["../src/index.ts"],"sourcesContent":["// @guidekit/vad — Silero VAD ONNX model wrapper for voice activity detection\nimport * as ort from 'onnxruntime-web';\n\nexport const VAD_VERSION = '0.1.0-beta.2';\n\n// ---------------------------------------------------------------------------\n// Constants\n// ---------------------------------------------------------------------------\n\nconst LOG_PREFIX = '[GuideKit:VAD]';\n\n/** Default CDN URL for the Silero VAD ONNX model (v5). */\nconst DEFAULT_MODEL_URL =\n  'https://cdn.jsdelivr.net/npm/@ricky0123/vad-web@0.0.20/dist/silero_vad_v5.onnx';\n\n/** Cache API key used for persisting the downloaded ONNX model. */\nconst CACHE_NAME = `guidekit-vad-v${VAD_VERSION}`;\nconst CACHE_MODEL_KEY = 'model.onnx';\n\n/** Silero VAD frame size: 512 samples at 16 kHz = 32 ms per frame. */\nconst FRAME_SIZE = 512;\n\n/** Target sample rate for VAD processing. */\nconst TARGET_SAMPLE_RATE = 16000;\n\n/** Duration (in ms) of audio collected for noise floor calibration. */\nconst CALIBRATION_DURATION_MS = 500;\n\n/** Hidden/cell state size for Silero VAD v5 LSTM. */\nconst STATE_SIZE = 128;\n\n// ---------------------------------------------------------------------------\n// Types\n// ---------------------------------------------------------------------------\n\nexport interface VADOptions {\n  /** Speech probability threshold (0-1). Default: 0.5 */\n  threshold?: number;\n  /** Minimum speech duration in ms to trigger start. Default: 300 */\n  minSpeechDurationMs?: number;\n  /** Silence duration in ms after speech to trigger end. Default: 500 */\n  silenceDurationMs?: number;\n  /** Sample rate. Default: 16000 */\n  sampleRate?: number;\n  /** Enable debug logging. Default: false */\n  debug?: boolean;\n  /** Custom URL for the Silero VAD ONNX model file. */\n  modelUrl?: string;\n}\n\nexport interface VADEvent {\n  type: 'speech-start' | 'speech-end' | 'vad-ready';\n  timestamp: number;\n  /** Speech probability (0-1) at the moment of the event. */\n  probability?: number;\n}\n\ntype VADEventType = VADEvent['type'];\ntype VADCallback = (event: VADEvent) => void;\n\n// ---------------------------------------------------------------------------\n// Utility: Cache API helpers\n// ---------------------------------------------------------------------------\n\nasync function loadModelFromCache(): Promise<ArrayBuffer | null> {\n  if (typeof caches === 'undefined') return null;\n  try {\n    const cache = await caches.open(CACHE_NAME);\n    const response = await cache.match(CACHE_MODEL_KEY);\n    return response ? response.arrayBuffer() : null;\n  } catch {\n    // Cache API may be unavailable in certain contexts (e.g. opaque origins).\n    return null;\n  }\n}\n\nasync function saveModelToCache(data: ArrayBuffer): Promise<void> {\n  if (typeof caches === 'undefined') return;\n  try {\n    const cache = await caches.open(CACHE_NAME);\n    await cache.put(CACHE_MODEL_KEY, new Response(data));\n  } catch {\n    // Silently ignore cache write failures.\n  }\n}\n\n// ---------------------------------------------------------------------------\n// Utility: Resampler\n// ---------------------------------------------------------------------------\n\n/**\n * Simple linear-interpolation resampler from `inputRate` to `outputRate`.\n * Adequate for VAD where perceptual audio quality is irrelevant.\n */\nfunction resample(\n  input: Float32Array,\n  inputRate: number,\n  outputRate: number,\n): Float32Array {\n  if (inputRate === outputRate) return input;\n  const ratio = inputRate / outputRate;\n  const outputLength = Math.round(input.length / ratio);\n  const output = new Float32Array(outputLength);\n  for (let i = 0; i < outputLength; i++) {\n    const srcIndex = i * ratio;\n    const srcFloor = Math.floor(srcIndex);\n    const srcCeil = Math.min(srcFloor + 1, input.length - 1);\n    const frac = srcIndex - srcFloor;\n    output[i] = (input[srcFloor] as number) * (1 - frac) + (input[srcCeil] as number) * frac;\n  }\n  return output;\n}\n\n// ---------------------------------------------------------------------------\n// SileroVAD\n// ---------------------------------------------------------------------------\n\nexport class SileroVAD {\n  // Options (resolved with defaults)\n  private readonly _threshold: number;\n  private readonly _minSpeechDurationMs: number;\n  private readonly _silenceDurationMs: number;\n  private readonly _sampleRate: number;\n  private readonly _debug: boolean;\n  private readonly _modelUrl: string;\n\n  // ONNX Runtime state\n  private _session: ort.InferenceSession | null = null;\n  private _h: ort.Tensor | null = null;\n  private _c: ort.Tensor | null = null;\n\n  // Audio pipeline\n  private _audioContext: AudioContext | null = null;\n  private _ownsAudioContext = false;\n  private _sourceNode: MediaStreamAudioSourceNode | null = null;\n  private _workletNode: AudioWorkletNode | ScriptProcessorNode | null = null;\n  private _stream: MediaStream | null = null;\n\n  // Frame buffer for accumulating resampled samples into FRAME_SIZE chunks\n  private _frameBuffer: Float32Array = new Float32Array(0);\n  private _frameBufferOffset = 0;\n\n  // State tracking\n  private _isReady = false;\n  private _isSpeaking = false;\n  private _isStarted = false;\n  private _isDestroyed = false;\n\n  // Duration tracking (in frames)\n  private _consecutiveSpeechFrames = 0;\n  private _consecutiveSilenceFrames = 0;\n  private _frameDurationMs: number;\n  private _minSpeechFrames: number;\n  private _silenceFrames: number;\n\n  // Noise floor calibration\n  private _isCalibrating = false;\n  private _calibrationSamples: number[] = [];\n  private _calibrationFramesNeeded = 0;\n  private _calibratedThreshold: number;\n\n  // Event listeners\n  private _listeners: Map<VADEventType, Set<VADCallback>> = new Map();\n\n  // Processing lock to serialise frame inference\n  private _processingPromise: Promise<void> = Promise.resolve();\n\n  constructor(options?: VADOptions) {\n    this._threshold = options?.threshold ?? 0.5;\n    this._minSpeechDurationMs = options?.minSpeechDurationMs ?? 300;\n    this._silenceDurationMs = options?.silenceDurationMs ?? 500;\n    this._sampleRate = options?.sampleRate ?? TARGET_SAMPLE_RATE;\n    this._debug = options?.debug ?? false;\n    this._modelUrl = options?.modelUrl ?? DEFAULT_MODEL_URL;\n    this._calibratedThreshold = this._threshold;\n\n    // Pre-compute frame-duration-based counters\n    this._frameDurationMs = (FRAME_SIZE / this._sampleRate) * 1000;\n    this._minSpeechFrames = Math.ceil(this._minSpeechDurationMs / this._frameDurationMs);\n    this._silenceFrames = Math.ceil(this._silenceDurationMs / this._frameDurationMs);\n\n    this._log('Created with options', {\n      threshold: this._threshold,\n      minSpeechDurationMs: this._minSpeechDurationMs,\n      silenceDurationMs: this._silenceDurationMs,\n      sampleRate: this._sampleRate,\n      modelUrl: this._modelUrl,\n    });\n  }\n\n  // -------------------------------------------------------------------------\n  // Public API\n  // -------------------------------------------------------------------------\n\n  /** Load the ONNX model. Uses Cache API for persistence across sessions. */\n  async init(): Promise<void> {\n    if (this._isDestroyed) {\n      throw new Error(`${LOG_PREFIX} Cannot init after destroy`);\n    }\n    if (this._isReady) {\n      this._log('Already initialised — skipping');\n      return;\n    }\n\n    this._log('Initialising...');\n\n    // 1. Attempt to load model bytes from cache, falling back to network.\n    let modelBuffer = await loadModelFromCache();\n    if (modelBuffer) {\n      this._log('Loaded model from Cache API');\n    } else {\n      this._log('Fetching model from', this._modelUrl);\n      const response = await fetch(this._modelUrl);\n      if (!response.ok) {\n        throw new Error(\n          `${LOG_PREFIX} Failed to fetch model: ${response.status} ${response.statusText}`,\n        );\n      }\n      modelBuffer = await response.arrayBuffer();\n      this._log('Model fetched, size:', modelBuffer.byteLength, 'bytes');\n\n      // Persist to Cache API for next time.\n      await saveModelToCache(modelBuffer);\n      this._log('Model saved to Cache API');\n    }\n\n    // 2. Create ONNX InferenceSession.\n    this._session = await ort.InferenceSession.create(modelBuffer, {\n      executionProviders: ['wasm'],\n      graphOptimizationLevel: 'all',\n    });\n\n    // 3. Initialise LSTM hidden/cell state tensors (zeros).\n    this._resetStates();\n\n    this._isReady = true;\n    this._log('Model loaded and ready');\n\n    this._emit({\n      type: 'vad-ready',\n      timestamp: Date.now(),\n    });\n  }\n\n  /**\n   * Process a single audio frame (512 samples at 16 kHz).\n   * Returns the speech probability (0-1).\n   */\n  async processFrame(audioData: Float32Array): Promise<number> {\n    if (!this._session) {\n      throw new Error(`${LOG_PREFIX} Model not loaded. Call init() first.`);\n    }\n    if (audioData.length !== FRAME_SIZE) {\n      throw new Error(\n        `${LOG_PREFIX} Expected ${FRAME_SIZE} samples, got ${audioData.length}`,\n      );\n    }\n\n    const inputTensor = new ort.Tensor('float32', audioData, [1, FRAME_SIZE]);\n    const srTensor = new ort.Tensor('int64', BigInt64Array.from([BigInt(this._sampleRate)]), [1]);\n\n    const feeds: Record<string, ort.Tensor> = {\n      input: inputTensor,\n      sr: srTensor,\n      h: this._h!,\n      c: this._c!,\n    };\n\n    const results = await this._session.run(feeds);\n\n    // Update LSTM hidden/cell states for the next frame.\n    this._h = results['hn'] as ort.Tensor;\n    this._c = results['cn'] as ort.Tensor;\n\n    const probability = (results['output'] as ort.Tensor).data[0] as number;\n    return probability;\n  }\n\n  /** Start VAD processing on a MediaStream (typically from getUserMedia). */\n  start(stream: MediaStream): void {\n    if (this._isDestroyed) {\n      throw new Error(`${LOG_PREFIX} Cannot start after destroy`);\n    }\n    if (!this._isReady) {\n      throw new Error(`${LOG_PREFIX} Model not loaded. Call init() first.`);\n    }\n    if (this._isStarted) {\n      this._log('Already started — stopping previous session first');\n      this.stop();\n    }\n\n    this._log('Starting VAD on MediaStream');\n    this._stream = stream;\n    this._isStarted = true;\n\n    // Reset speech tracking state.\n    this._isSpeaking = false;\n    this._consecutiveSpeechFrames = 0;\n    this._consecutiveSilenceFrames = 0;\n    this._frameBuffer = new Float32Array(FRAME_SIZE);\n    this._frameBufferOffset = 0;\n\n    // Reset LSTM states for a fresh stream.\n    this._resetStates();\n\n    // Begin noise floor calibration.\n    this._isCalibrating = true;\n    this._calibrationSamples = [];\n    this._calibrationFramesNeeded = Math.max(1, Math.floor(\n      (CALIBRATION_DURATION_MS / 1000) * this._sampleRate / FRAME_SIZE,\n    ));\n    this._log('Calibrating noise floor for', this._calibrationFramesNeeded, 'frames');\n\n    // Build the audio processing pipeline.\n    this._setupAudioPipeline(stream);\n  }\n\n  /** Stop VAD processing and release audio resources (but keep the model). */\n  stop(): void {\n    if (!this._isStarted) return;\n\n    this._log('Stopping VAD');\n\n    // Tear down audio nodes.\n    this._teardownAudioPipeline();\n\n    // If we were speaking, emit speech-end.\n    if (this._isSpeaking) {\n      this._isSpeaking = false;\n      this._emit({\n        type: 'speech-end',\n        timestamp: Date.now(),\n        probability: 0,\n      });\n    }\n\n    // Reset state.\n    this._isStarted = false;\n    this._isSpeaking = false;\n    this._consecutiveSpeechFrames = 0;\n    this._consecutiveSilenceFrames = 0;\n    this._frameBufferOffset = 0;\n    this._isCalibrating = false;\n    this._calibrationSamples = [];\n    this._processingPromise = Promise.resolve();\n\n    this._log('VAD stopped');\n  }\n\n  /** Register a callback for speech-start events. Returns an unsubscribe function. */\n  onSpeechStart(callback: VADCallback): () => void {\n    return this._on('speech-start', callback);\n  }\n\n  /** Register a callback for speech-end events. Returns an unsubscribe function. */\n  onSpeechEnd(callback: VADCallback): () => void {\n    return this._on('speech-end', callback);\n  }\n\n  /** Register a callback for vad-ready events. Returns an unsubscribe function. */\n  onReady(callback: VADCallback): () => void {\n    return this._on('vad-ready', callback);\n  }\n\n  /** Whether the ONNX model is loaded and ready. */\n  get isReady(): boolean {\n    return this._isReady;\n  }\n\n  /** Whether speech is currently detected. */\n  get isSpeaking(): boolean {\n    return this._isSpeaking;\n  }\n\n  /** The MediaStream currently being processed, or null. */\n  get stream(): MediaStream | null {\n    return this._stream;\n  }\n\n  /** Release ONNX model session and all audio resources. */\n  async destroy(): Promise<void> {\n    if (this._isDestroyed) return;\n    this._log('Destroying...');\n\n    this.stop();\n\n    if (this._session) {\n      await this._session.release();\n      this._session = null;\n    }\n\n    // Dispose tensors.\n    this._h?.dispose();\n    this._c?.dispose();\n    this._h = null;\n    this._c = null;\n\n    this._isReady = false;\n    this._isDestroyed = true;\n    this._listeners.clear();\n\n    this._log('Destroyed');\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: Event system\n  // -------------------------------------------------------------------------\n\n  private _on(type: VADEventType, callback: VADCallback): () => void {\n    let set = this._listeners.get(type);\n    if (!set) {\n      set = new Set();\n      this._listeners.set(type, set);\n    }\n    set.add(callback);\n    return () => {\n      set!.delete(callback);\n    };\n  }\n\n  private _emit(event: VADEvent): void {\n    const set = this._listeners.get(event.type);\n    if (!set) return;\n    for (const cb of set) {\n      try {\n        cb(event);\n      } catch (err) {\n        // eslint-disable-next-line no-console\n        console.error(`${LOG_PREFIX} Error in ${event.type} callback:`, err);\n      }\n    }\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: Audio pipeline\n  // -------------------------------------------------------------------------\n\n  private _setupAudioPipeline(stream: MediaStream): void {\n    // Determine the incoming sample rate.\n    const tracks = stream.getAudioTracks();\n    const trackSettings = tracks[0]?.getSettings();\n    const inputSampleRate = trackSettings?.sampleRate ?? 48000;\n\n    this._log('Input sample rate:', inputSampleRate);\n\n    // Create AudioContext at the input sample rate so we don't double-resample.\n    // SSR guard: AudioContext may not exist.\n    if (typeof AudioContext === 'undefined' && typeof webkitAudioContext === 'undefined') {\n      throw new Error(`${LOG_PREFIX} AudioContext is not available in this environment`);\n    }\n\n    const AudioContextClass =\n      typeof AudioContext !== 'undefined'\n        ? AudioContext\n        : // eslint-disable-next-line @typescript-eslint/no-explicit-any\n          (globalThis as any).webkitAudioContext as typeof AudioContext;\n\n    this._audioContext = new AudioContextClass({ sampleRate: inputSampleRate });\n    this._ownsAudioContext = true;\n\n    this._sourceNode = this._audioContext.createMediaStreamSource(stream);\n\n    // Try AudioWorklet first, fall back to ScriptProcessorNode.\n    this._setupScriptProcessor(inputSampleRate);\n  }\n\n  /**\n   * ScriptProcessorNode fallback (works everywhere, including Safari).\n   * We use a buffer size of 4096 which gives ~85 ms of audio at 48 kHz.\n   */\n  private _setupScriptProcessor(inputSampleRate: number): void {\n    if (!this._audioContext || !this._sourceNode) return;\n\n    // Buffer size must be a power of 2: 256, 512, 1024, 2048, 4096, 8192, 16384.\n    const bufferSize = 4096;\n    const processor = this._audioContext.createScriptProcessor(bufferSize, 1, 1);\n\n    processor.onaudioprocess = (event: AudioProcessingEvent) => {\n      if (!this._isStarted) return;\n\n      const inputData = event.inputBuffer.getChannelData(0);\n\n      // Resample to target rate if needed.\n      const resampled =\n        inputSampleRate !== this._sampleRate\n          ? resample(inputData, inputSampleRate, this._sampleRate)\n          : new Float32Array(inputData);\n\n      // Feed resampled audio into frame-sized chunks.\n      this._feedAudio(resampled);\n    };\n\n    this._sourceNode.connect(processor);\n    processor.connect(this._audioContext.destination);\n    this._workletNode = processor;\n\n    this._log('Audio pipeline set up (ScriptProcessorNode)');\n  }\n\n  /**\n   * Accumulate resampled audio into FRAME_SIZE chunks and process each full frame.\n   */\n  private _feedAudio(samples: Float32Array): void {\n    let offset = 0;\n\n    while (offset < samples.length) {\n      const remaining = FRAME_SIZE - this._frameBufferOffset;\n      const available = samples.length - offset;\n      const toCopy = Math.min(remaining, available);\n\n      this._frameBuffer.set(\n        samples.subarray(offset, offset + toCopy),\n        this._frameBufferOffset,\n      );\n      this._frameBufferOffset += toCopy;\n      offset += toCopy;\n\n      if (this._frameBufferOffset === FRAME_SIZE) {\n        const frame = new Float32Array(this._frameBuffer);\n        this._frameBufferOffset = 0;\n\n        // Serialise inference calls to avoid overlapping ONNX sessions.\n        this._processingPromise = this._processingPromise.then(() =>\n          this._handleFrame(frame),\n        );\n      }\n    }\n  }\n\n  /**\n   * Process a single FRAME_SIZE frame: run inference and update speech state.\n   */\n  private async _handleFrame(frame: Float32Array): Promise<void> {\n    if (!this._isStarted || !this._session) return;\n\n    let probability: number;\n    try {\n      probability = await this.processFrame(frame);\n    } catch (err) {\n      if (this._debug) {\n        // eslint-disable-next-line no-console\n        console.error(`${LOG_PREFIX} Inference error:`, err);\n      }\n      return;\n    }\n\n    // Noise floor calibration phase.\n    if (this._isCalibrating) {\n      this._calibrationSamples.push(probability);\n\n      if (this._calibrationSamples.length >= this._calibrationFramesNeeded) {\n        this._finishCalibration();\n      }\n      return;\n    }\n\n    // Speech state machine.\n    const isSpeechFrame = probability >= this._calibratedThreshold;\n\n    if (isSpeechFrame) {\n      this._consecutiveSpeechFrames++;\n      this._consecutiveSilenceFrames = 0;\n\n      if (!this._isSpeaking && this._consecutiveSpeechFrames >= this._minSpeechFrames) {\n        this._isSpeaking = true;\n        this._log('Speech started, probability:', probability.toFixed(3));\n        this._emit({\n          type: 'speech-start',\n          timestamp: Date.now(),\n          probability,\n        });\n      }\n    } else {\n      this._consecutiveSilenceFrames++;\n      // Do NOT reset _consecutiveSpeechFrames here — only reset when speech-end fires.\n\n      if (this._isSpeaking && this._consecutiveSilenceFrames >= this._silenceFrames) {\n        this._isSpeaking = false;\n        this._consecutiveSpeechFrames = 0;\n        this._log('Speech ended, probability:', probability.toFixed(3));\n        this._emit({\n          type: 'speech-end',\n          timestamp: Date.now(),\n          probability,\n        });\n      }\n    }\n  }\n\n  private _finishCalibration(): void {\n    if (this._calibrationSamples.length === 0) {\n      this._isCalibrating = false;\n      return;\n    }\n\n    // Compute average noise floor probability.\n    const sum = this._calibrationSamples.reduce((a, b) => a + b, 0);\n    const avgNoise = sum / this._calibrationSamples.length;\n\n    // If the ambient noise floor is high, nudge the threshold above it.\n    // We add a margin so we don't constantly trigger on background noise.\n    const NOISE_MARGIN = 0.15;\n    if (avgNoise + NOISE_MARGIN > this._threshold) {\n      this._calibratedThreshold = Math.min(avgNoise + NOISE_MARGIN, 0.95);\n      this._log(\n        'Noise floor is high. Adjusted threshold from',\n        this._threshold.toFixed(3),\n        'to',\n        this._calibratedThreshold.toFixed(3),\n        '(avg noise:',\n        avgNoise.toFixed(3) + ')',\n      );\n    } else {\n      this._calibratedThreshold = this._threshold;\n      this._log('Noise floor OK, avg:', avgNoise.toFixed(3), '— keeping threshold at', this._threshold.toFixed(3));\n    }\n\n    // Recompute frame counters in case threshold changed min speech behaviour.\n    this._minSpeechFrames = Math.ceil(this._minSpeechDurationMs / this._frameDurationMs);\n    this._silenceFrames = Math.ceil(this._silenceDurationMs / this._frameDurationMs);\n\n    this._isCalibrating = false;\n    this._calibrationSamples = [];\n  }\n\n  private _teardownAudioPipeline(): void {\n    if (this._workletNode) {\n      try {\n        this._workletNode.disconnect();\n      } catch {\n        // Ignore disconnect errors.\n      }\n      if ('onaudioprocess' in this._workletNode) {\n        (this._workletNode as ScriptProcessorNode).onaudioprocess = null;\n      }\n      this._workletNode = null;\n    }\n\n    if (this._sourceNode) {\n      try {\n        this._sourceNode.disconnect();\n      } catch {\n        // Ignore.\n      }\n      this._sourceNode = null;\n    }\n\n    if (this._audioContext && this._ownsAudioContext) {\n      try {\n        void this._audioContext.close();\n      } catch {\n        // Ignore.\n      }\n      this._audioContext = null;\n      this._ownsAudioContext = false;\n    }\n\n    this._stream = null;\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: ONNX state helpers\n  // -------------------------------------------------------------------------\n\n  /** Reset the LSTM hidden and cell states to zeros. */\n  private _resetStates(): void {\n    // Dispose any existing tensors to free memory.\n    this._h?.dispose();\n    this._c?.dispose();\n\n    const zeros = new Float32Array(2 * STATE_SIZE).fill(0);\n    this._h = new ort.Tensor('float32', zeros.slice(0, STATE_SIZE), [2, 1, 64]);\n    this._c = new ort.Tensor('float32', zeros.slice(STATE_SIZE), [2, 1, 64]);\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: Logging\n  // -------------------------------------------------------------------------\n\n  private _log(...args: unknown[]): void {\n    if (!this._debug) return;\n    // eslint-disable-next-line no-console\n    console.log(LOG_PREFIX, ...args);\n  }\n}\n\n// ---------------------------------------------------------------------------\n// Convenience factory\n// ---------------------------------------------------------------------------\n\n/**\n * Create and initialise a SileroVAD instance in one call.\n *\n * ```ts\n * const vad = await createVAD({ debug: true });\n * vad.onSpeechStart(() => console.log('speaking'));\n * vad.start(stream);\n * ```\n */\nexport async function createVAD(options?: VADOptions): Promise<SileroVAD> {\n  const vad = new SileroVAD(options);\n  await vad.init();\n  return vad;\n}\n\n// Re-export the frame size constant so consumers can align their buffers.\nexport { FRAME_SIZE, TARGET_SAMPLE_RATE };\n\n// Type-only declaration for environments that provide webkitAudioContext.\ndeclare global {\n  // eslint-disable-next-line no-var\n  var webkitAudioContext: typeof AudioContext | undefined;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AACA,UAAqB;AAEd,IAAM,cAAc;AAM3B,IAAM,aAAa;AAGnB,IAAM,oBACJ;AAGF,IAAM,aAAa,iBAAiB,WAAW;AAC/C,IAAM,kBAAkB;AAGxB,IAAM,aAAa;AAGnB,IAAM,qBAAqB;AAG3B,IAAM,0BAA0B;AAGhC,IAAM,aAAa;AAmCnB,eAAe,qBAAkD;AAC/D,MAAI,OAAO,WAAW,YAAa,QAAO;AAC1C,MAAI;AACF,UAAM,QAAQ,MAAM,OAAO,KAAK,UAAU;AAC1C,UAAM,WAAW,MAAM,MAAM,MAAM,eAAe;AAClD,WAAO,WAAW,SAAS,YAAY,IAAI;AAAA,EAC7C,QAAQ;AAEN,WAAO;AAAA,EACT;AACF;AAEA,eAAe,iBAAiB,MAAkC;AAChE,MAAI,OAAO,WAAW,YAAa;AACnC,MAAI;AACF,UAAM,QAAQ,MAAM,OAAO,KAAK,UAAU;AAC1C,UAAM,MAAM,IAAI,iBAAiB,IAAI,SAAS,IAAI,CAAC;AAAA,EACrD,QAAQ;AAAA,EAER;AACF;AAUA,SAAS,SACP,OACA,WACA,YACc;AACd,MAAI,cAAc,WAAY,QAAO;AACrC,QAAM,QAAQ,YAAY;AAC1B,QAAM,eAAe,KAAK,MAAM,MAAM,SAAS,KAAK;AACpD,QAAM,SAAS,IAAI,aAAa,YAAY;AAC5C,WAAS,IAAI,GAAG,IAAI,cAAc,KAAK;AACrC,UAAM,WAAW,IAAI;AACrB,UAAM,WAAW,KAAK,MAAM,QAAQ;AACpC,UAAM,UAAU,KAAK,IAAI,WAAW,GAAG,MAAM,SAAS,CAAC;AACvD,UAAM,OAAO,WAAW;AACxB,WAAO,CAAC,IAAK,MAAM,QAAQ,KAAgB,IAAI,QAAS,MAAM,OAAO,IAAe;AAAA,EACtF;AACA,SAAO;AACT;AAMO,IAAM,YAAN,MAAgB;AAAA;AAAA,EAEJ;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA;AAAA,EAGT,WAAwC;AAAA,EACxC,KAAwB;AAAA,EACxB,KAAwB;AAAA;AAAA,EAGxB,gBAAqC;AAAA,EACrC,oBAAoB;AAAA,EACpB,cAAiD;AAAA,EACjD,eAA8D;AAAA,EAC9D,UAA8B;AAAA;AAAA,EAG9B,eAA6B,IAAI,aAAa,CAAC;AAAA,EAC/C,qBAAqB;AAAA;AAAA,EAGrB,WAAW;AAAA,EACX,cAAc;AAAA,EACd,aAAa;AAAA,EACb,eAAe;AAAA;AAAA,EAGf,2BAA2B;AAAA,EAC3B,4BAA4B;AAAA,EAC5B;AAAA,EACA;AAAA,EACA;AAAA;AAAA,EAGA,iBAAiB;AAAA,EACjB,sBAAgC,CAAC;AAAA,EACjC,2BAA2B;AAAA,EAC3B;AAAA;AAAA,EAGA,aAAkD,oBAAI,IAAI;AAAA;AAAA,EAG1D,qBAAoC,QAAQ,QAAQ;AAAA,EAE5D,YAAY,SAAsB;AAChC,SAAK,aAAa,SAAS,aAAa;AACxC,SAAK,uBAAuB,SAAS,uBAAuB;AAC5D,SAAK,qBAAqB,SAAS,qBAAqB;AACxD,SAAK,cAAc,SAAS,cAAc;AAC1C,SAAK,SAAS,SAAS,SAAS;AAChC,SAAK,YAAY,SAAS,YAAY;AACtC,SAAK,uBAAuB,KAAK;AAGjC,SAAK,mBAAoB,aAAa,KAAK,cAAe;AAC1D,SAAK,mBAAmB,KAAK,KAAK,KAAK,uBAAuB,KAAK,gBAAgB;AACnF,SAAK,iBAAiB,KAAK,KAAK,KAAK,qBAAqB,KAAK,gBAAgB;AAE/E,SAAK,KAAK,wBAAwB;AAAA,MAChC,WAAW,KAAK;AAAA,MAChB,qBAAqB,KAAK;AAAA,MAC1B,mBAAmB,KAAK;AAAA,MACxB,YAAY,KAAK;AAAA,MACjB,UAAU,KAAK;AAAA,IACjB,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,OAAsB;AAC1B,QAAI,KAAK,cAAc;AACrB,YAAM,IAAI,MAAM,GAAG,UAAU,4BAA4B;AAAA,IAC3D;AACA,QAAI,KAAK,UAAU;AACjB,WAAK,KAAK,qCAAgC;AAC1C;AAAA,IACF;AAEA,SAAK,KAAK,iBAAiB;AAG3B,QAAI,cAAc,MAAM,mBAAmB;AAC3C,QAAI,aAAa;AACf,WAAK,KAAK,6BAA6B;AAAA,IACzC,OAAO;AACL,WAAK,KAAK,uBAAuB,KAAK,SAAS;AAC/C,YAAM,WAAW,MAAM,MAAM,KAAK,SAAS;AAC3C,UAAI,CAAC,SAAS,IAAI;AAChB,cAAM,IAAI;AAAA,UACR,GAAG,UAAU,2BAA2B,SAAS,MAAM,IAAI,SAAS,UAAU;AAAA,QAChF;AAAA,MACF;AACA,oBAAc,MAAM,SAAS,YAAY;AACzC,WAAK,KAAK,wBAAwB,YAAY,YAAY,OAAO;AAGjE,YAAM,iBAAiB,WAAW;AAClC,WAAK,KAAK,0BAA0B;AAAA,IACtC;AAGA,SAAK,WAAW,MAAU,qBAAiB,OAAO,aAAa;AAAA,MAC7D,oBAAoB,CAAC,MAAM;AAAA,MAC3B,wBAAwB;AAAA,IAC1B,CAAC;AAGD,SAAK,aAAa;AAElB,SAAK,WAAW;AAChB,SAAK,KAAK,wBAAwB;AAElC,SAAK,MAAM;AAAA,MACT,MAAM;AAAA,MACN,WAAW,KAAK,IAAI;AAAA,IACtB,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,aAAa,WAA0C;AAC3D,QAAI,CAAC,KAAK,UAAU;AAClB,YAAM,IAAI,MAAM,GAAG,UAAU,uCAAuC;AAAA,IACtE;AACA,QAAI,UAAU,WAAW,YAAY;AACnC,YAAM,IAAI;AAAA,QACR,GAAG,UAAU,aAAa,UAAU,iBAAiB,UAAU,MAAM;AAAA,MACvE;AAAA,IACF;AAEA,UAAM,cAAc,IAAQ,WAAO,WAAW,WAAW,CAAC,GAAG,UAAU,CAAC;AACxE,UAAM,WAAW,IAAQ,WAAO,SAAS,cAAc,KAAK,CAAC,OAAO,KAAK,WAAW,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;AAE5F,UAAM,QAAoC;AAAA,MACxC,OAAO;AAAA,MACP,IAAI;AAAA,MACJ,GAAG,KAAK;AAAA,MACR,GAAG,KAAK;AAAA,IACV;AAEA,UAAM,UAAU,MAAM,KAAK,SAAS,IAAI,KAAK;AAG7C,SAAK,KAAK,QAAQ,IAAI;AACtB,SAAK,KAAK,QAAQ,IAAI;AAEtB,UAAM,cAAe,QAAQ,QAAQ,EAAiB,KAAK,CAAC;AAC5D,WAAO;AAAA,EACT;AAAA;AAAA,EAGA,MAAM,QAA2B;AAC/B,QAAI,KAAK,cAAc;AACrB,YAAM,IAAI,MAAM,GAAG,UAAU,6BAA6B;AAAA,IAC5D;AACA,QAAI,CAAC,KAAK,UAAU;AAClB,YAAM,IAAI,MAAM,GAAG,UAAU,uCAAuC;AAAA,IACtE;AACA,QAAI,KAAK,YAAY;AACnB,WAAK,KAAK,wDAAmD;AAC7D,WAAK,KAAK;AAAA,IACZ;AAEA,SAAK,KAAK,6BAA6B;AACvC,SAAK,UAAU;AACf,SAAK,aAAa;AAGlB,SAAK,cAAc;AACnB,SAAK,2BAA2B;AAChC,SAAK,4BAA4B;AACjC,SAAK,eAAe,IAAI,aAAa,UAAU;AAC/C,SAAK,qBAAqB;AAG1B,SAAK,aAAa;AAGlB,SAAK,iBAAiB;AACtB,SAAK,sBAAsB,CAAC;AAC5B,SAAK,2BAA2B,KAAK,IAAI,GAAG,KAAK;AAAA,MAC9C,0BAA0B,MAAQ,KAAK,cAAc;AAAA,IACxD,CAAC;AACD,SAAK,KAAK,+BAA+B,KAAK,0BAA0B,QAAQ;AAGhF,SAAK,oBAAoB,MAAM;AAAA,EACjC;AAAA;AAAA,EAGA,OAAa;AACX,QAAI,CAAC,KAAK,WAAY;AAEtB,SAAK,KAAK,cAAc;AAGxB,SAAK,uBAAuB;AAG5B,QAAI,KAAK,aAAa;AACpB,WAAK,cAAc;AACnB,WAAK,MAAM;AAAA,QACT,MAAM;AAAA,QACN,WAAW,KAAK,IAAI;AAAA,QACpB,aAAa;AAAA,MACf,CAAC;AAAA,IACH;AAGA,SAAK,aAAa;AAClB,SAAK,cAAc;AACnB,SAAK,2BAA2B;AAChC,SAAK,4BAA4B;AACjC,SAAK,qBAAqB;AAC1B,SAAK,iBAAiB;AACtB,SAAK,sBAAsB,CAAC;AAC5B,SAAK,qBAAqB,QAAQ,QAAQ;AAE1C,SAAK,KAAK,aAAa;AAAA,EACzB;AAAA;AAAA,EAGA,cAAc,UAAmC;AAC/C,WAAO,KAAK,IAAI,gBAAgB,QAAQ;AAAA,EAC1C;AAAA;AAAA,EAGA,YAAY,UAAmC;AAC7C,WAAO,KAAK,IAAI,cAAc,QAAQ;AAAA,EACxC;AAAA;AAAA,EAGA,QAAQ,UAAmC;AACzC,WAAO,KAAK,IAAI,aAAa,QAAQ;AAAA,EACvC;AAAA;AAAA,EAGA,IAAI,UAAmB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,aAAsB;AACxB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,SAA6B;AAC/B,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,MAAM,UAAyB;AAC7B,QAAI,KAAK,aAAc;AACvB,SAAK,KAAK,eAAe;AAEzB,SAAK,KAAK;AAEV,QAAI,KAAK,UAAU;AACjB,YAAM,KAAK,SAAS,QAAQ;AAC5B,WAAK,WAAW;AAAA,IAClB;AAGA,SAAK,IAAI,QAAQ;AACjB,SAAK,IAAI,QAAQ;AACjB,SAAK,KAAK;AACV,SAAK,KAAK;AAEV,SAAK,WAAW;AAChB,SAAK,eAAe;AACpB,SAAK,WAAW,MAAM;AAEtB,SAAK,KAAK,WAAW;AAAA,EACvB;AAAA;AAAA;AAAA;AAAA,EAMQ,IAAI,MAAoB,UAAmC;AACjE,QAAI,MAAM,KAAK,WAAW,IAAI,IAAI;AAClC,QAAI,CAAC,KAAK;AACR,YAAM,oBAAI,IAAI;AACd,WAAK,WAAW,IAAI,MAAM,GAAG;AAAA,IAC/B;AACA,QAAI,IAAI,QAAQ;AAChB,WAAO,MAAM;AACX,UAAK,OAAO,QAAQ;AAAA,IACtB;AAAA,EACF;AAAA,EAEQ,MAAM,OAAuB;AACnC,UAAM,MAAM,KAAK,WAAW,IAAI,MAAM,IAAI;AAC1C,QAAI,CAAC,IAAK;AACV,eAAW,MAAM,KAAK;AACpB,UAAI;AACF,WAAG,KAAK;AAAA,MACV,SAAS,KAAK;AAEZ,gBAAQ,MAAM,GAAG,UAAU,aAAa,MAAM,IAAI,cAAc,GAAG;AAAA,MACrE;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAMQ,oBAAoB,QAA2B;AAErD,UAAM,SAAS,OAAO,eAAe;AACrC,UAAM,gBAAgB,OAAO,CAAC,GAAG,YAAY;AAC7C,UAAM,kBAAkB,eAAe,cAAc;AAErD,SAAK,KAAK,sBAAsB,eAAe;AAI/C,QAAI,OAAO,iBAAiB,eAAe,OAAO,uBAAuB,aAAa;AACpF,YAAM,IAAI,MAAM,GAAG,UAAU,oDAAoD;AAAA,IACnF;AAEA,UAAM,oBACJ,OAAO,iBAAiB,cACpB;AAAA;AAAA,MAEC,WAAmB;AAAA;AAE1B,SAAK,gBAAgB,IAAI,kBAAkB,EAAE,YAAY,gBAAgB,CAAC;AAC1E,SAAK,oBAAoB;AAEzB,SAAK,cAAc,KAAK,cAAc,wBAAwB,MAAM;AAGpE,SAAK,sBAAsB,eAAe;AAAA,EAC5C;AAAA;AAAA;AAAA;AAAA;AAAA,EAMQ,sBAAsB,iBAA+B;AAC3D,QAAI,CAAC,KAAK,iBAAiB,CAAC,KAAK,YAAa;AAG9C,UAAM,aAAa;AACnB,UAAM,YAAY,KAAK,cAAc,sBAAsB,YAAY,GAAG,CAAC;AAE3E,cAAU,iBAAiB,CAAC,UAAgC;AAC1D,UAAI,CAAC,KAAK,WAAY;AAEtB,YAAM,YAAY,MAAM,YAAY,eAAe,CAAC;AAGpD,YAAM,YACJ,oBAAoB,KAAK,cACrB,SAAS,WAAW,iBAAiB,KAAK,WAAW,IACrD,IAAI,aAAa,SAAS;AAGhC,WAAK,WAAW,SAAS;AAAA,IAC3B;AAEA,SAAK,YAAY,QAAQ,SAAS;AAClC,cAAU,QAAQ,KAAK,cAAc,WAAW;AAChD,SAAK,eAAe;AAEpB,SAAK,KAAK,6CAA6C;AAAA,EACzD;AAAA;AAAA;AAAA;AAAA,EAKQ,WAAW,SAA6B;AAC9C,QAAI,SAAS;AAEb,WAAO,SAAS,QAAQ,QAAQ;AAC9B,YAAM,YAAY,aAAa,KAAK;AACpC,YAAM,YAAY,QAAQ,SAAS;AACnC,YAAM,SAAS,KAAK,IAAI,WAAW,SAAS;AAE5C,WAAK,aAAa;AAAA,QAChB,QAAQ,SAAS,QAAQ,SAAS,MAAM;AAAA,QACxC,KAAK;AAAA,MACP;AACA,WAAK,sBAAsB;AAC3B,gBAAU;AAEV,UAAI,KAAK,uBAAuB,YAAY;AAC1C,cAAM,QAAQ,IAAI,aAAa,KAAK,YAAY;AAChD,aAAK,qBAAqB;AAG1B,aAAK,qBAAqB,KAAK,mBAAmB;AAAA,UAAK,MACrD,KAAK,aAAa,KAAK;AAAA,QACzB;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,MAAc,aAAa,OAAoC;AAC7D,QAAI,CAAC,KAAK,cAAc,CAAC,KAAK,SAAU;AAExC,QAAI;AACJ,QAAI;AACF,oBAAc,MAAM,KAAK,aAAa,KAAK;AAAA,IAC7C,SAAS,KAAK;AACZ,UAAI,KAAK,QAAQ;AAEf,gBAAQ,MAAM,GAAG,UAAU,qBAAqB,GAAG;AAAA,MACrD;AACA;AAAA,IACF;AAGA,QAAI,KAAK,gBAAgB;AACvB,WAAK,oBAAoB,KAAK,WAAW;AAEzC,UAAI,KAAK,oBAAoB,UAAU,KAAK,0BAA0B;AACpE,aAAK,mBAAmB;AAAA,MAC1B;AACA;AAAA,IACF;AAGA,UAAM,gBAAgB,eAAe,KAAK;AAE1C,QAAI,eAAe;AACjB,WAAK;AACL,WAAK,4BAA4B;AAEjC,UAAI,CAAC,KAAK,eAAe,KAAK,4BAA4B,KAAK,kBAAkB;AAC/E,aAAK,cAAc;AACnB,aAAK,KAAK,gCAAgC,YAAY,QAAQ,CAAC,CAAC;AAChE,aAAK,MAAM;AAAA,UACT,MAAM;AAAA,UACN,WAAW,KAAK,IAAI;AAAA,UACpB;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF,OAAO;AACL,WAAK;AAGL,UAAI,KAAK,eAAe,KAAK,6BAA6B,KAAK,gBAAgB;AAC7E,aAAK,cAAc;AACnB,aAAK,2BAA2B;AAChC,aAAK,KAAK,8BAA8B,YAAY,QAAQ,CAAC,CAAC;AAC9D,aAAK,MAAM;AAAA,UACT,MAAM;AAAA,UACN,WAAW,KAAK,IAAI;AAAA,UACpB;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,qBAA2B;AACjC,QAAI,KAAK,oBAAoB,WAAW,GAAG;AACzC,WAAK,iBAAiB;AACtB;AAAA,IACF;AAGA,UAAM,MAAM,KAAK,oBAAoB,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC9D,UAAM,WAAW,MAAM,KAAK,oBAAoB;AAIhD,UAAM,eAAe;AACrB,QAAI,WAAW,eAAe,KAAK,YAAY;AAC7C,WAAK,uBAAuB,KAAK,IAAI,WAAW,cAAc,IAAI;AAClE,WAAK;AAAA,QACH;AAAA,QACA,KAAK,WAAW,QAAQ,CAAC;AAAA,QACzB;AAAA,QACA,KAAK,qBAAqB,QAAQ,CAAC;AAAA,QACnC;AAAA,QACA,SAAS,QAAQ,CAAC,IAAI;AAAA,MACxB;AAAA,IACF,OAAO;AACL,WAAK,uBAAuB,KAAK;AACjC,WAAK,KAAK,wBAAwB,SAAS,QAAQ,CAAC,GAAG,+BAA0B,KAAK,WAAW,QAAQ,CAAC,CAAC;AAAA,IAC7G;AAGA,SAAK,mBAAmB,KAAK,KAAK,KAAK,uBAAuB,KAAK,gBAAgB;AACnF,SAAK,iBAAiB,KAAK,KAAK,KAAK,qBAAqB,KAAK,gBAAgB;AAE/E,SAAK,iBAAiB;AACtB,SAAK,sBAAsB,CAAC;AAAA,EAC9B;AAAA,EAEQ,yBAA+B;AACrC,QAAI,KAAK,cAAc;AACrB,UAAI;AACF,aAAK,aAAa,WAAW;AAAA,MAC/B,QAAQ;AAAA,MAER;AACA,UAAI,oBAAoB,KAAK,cAAc;AACzC,QAAC,KAAK,aAAqC,iBAAiB;AAAA,MAC9D;AACA,WAAK,eAAe;AAAA,IACtB;AAEA,QAAI,KAAK,aAAa;AACpB,UAAI;AACF,aAAK,YAAY,WAAW;AAAA,MAC9B,QAAQ;AAAA,MAER;AACA,WAAK,cAAc;AAAA,IACrB;AAEA,QAAI,KAAK,iBAAiB,KAAK,mBAAmB;AAChD,UAAI;AACF,aAAK,KAAK,cAAc,MAAM;AAAA,MAChC,QAAQ;AAAA,MAER;AACA,WAAK,gBAAgB;AACrB,WAAK,oBAAoB;AAAA,IAC3B;AAEA,SAAK,UAAU;AAAA,EACjB;AAAA;AAAA;AAAA;AAAA;AAAA,EAOQ,eAAqB;AAE3B,SAAK,IAAI,QAAQ;AACjB,SAAK,IAAI,QAAQ;AAEjB,UAAM,QAAQ,IAAI,aAAa,IAAI,UAAU,EAAE,KAAK,CAAC;AACrD,SAAK,KAAK,IAAQ,WAAO,WAAW,MAAM,MAAM,GAAG,UAAU,GAAG,CAAC,GAAG,GAAG,EAAE,CAAC;AAC1E,SAAK,KAAK,IAAQ,WAAO,WAAW,MAAM,MAAM,UAAU,GAAG,CAAC,GAAG,GAAG,EAAE,CAAC;AAAA,EACzE;AAAA;AAAA;AAAA;AAAA,EAMQ,QAAQ,MAAuB;AACrC,QAAI,CAAC,KAAK,OAAQ;AAElB,YAAQ,IAAI,YAAY,GAAG,IAAI;AAAA,EACjC;AACF;AAeA,eAAsB,UAAU,SAA0C;AACxE,QAAM,MAAM,IAAI,UAAU,OAAO;AACjC,QAAM,IAAI,KAAK;AACf,SAAO;AACT;","names":[]}

package/dist/index.d.cts

@@ -1,4 +1,4 @@
-declare const VAD_VERSION = "0.1.0";
+declare const VAD_VERSION = "0.1.0-beta.2";
 /** Silero VAD frame size: 512 samples at 16 kHz = 32 ms per frame. */
 declare const FRAME_SIZE = 512;
 /** Target sample rate for VAD processing. */

package/dist/index.d.ts

@@ -1,4 +1,4 @@
-declare const VAD_VERSION = "0.1.0";
+declare const VAD_VERSION = "0.1.0-beta.2";
 /** Silero VAD frame size: 512 samples at 16 kHz = 32 ms per frame. */
 declare const FRAME_SIZE = 512;
 /** Target sample rate for VAD processing. */

package/dist/index.js

@@ -1,6 +1,6 @@
 // src/index.ts
 import * as ort from "onnxruntime-web";
-var VAD_VERSION = "0.1.0";
+var VAD_VERSION = "0.1.0-beta.2";
 var LOG_PREFIX = "[GuideKit:VAD]";
 var DEFAULT_MODEL_URL = "https://cdn.jsdelivr.net/npm/@ricky0123/vad-web@0.0.20/dist/silero_vad_v5.onnx";
 var CACHE_NAME = `guidekit-vad-v${VAD_VERSION}`;
@@ -192,9 +192,9 @@ var SileroVAD = class {
     this._resetStates();
     this._isCalibrating = true;
     this._calibrationSamples = [];
-    this._calibrationFramesNeeded = Math.ceil(
+    this._calibrationFramesNeeded = Math.max(1, Math.floor(
       CALIBRATION_DURATION_MS / 1e3 * this._sampleRate / FRAME_SIZE
-    );
+    ));
     this._log("Calibrating noise floor for", this._calibrationFramesNeeded, "frames");
     this._setupAudioPipeline(stream);
   }

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/index.ts"],"sourcesContent":["// @guidekit/vad — Silero VAD ONNX model wrapper for voice activity detection\nimport * as ort from 'onnxruntime-web';\n\nexport const VAD_VERSION = '0.1.0';\n\n// ---------------------------------------------------------------------------\n// Constants\n// ---------------------------------------------------------------------------\n\nconst LOG_PREFIX = '[GuideKit:VAD]';\n\n/** Default CDN URL for the Silero VAD ONNX model (v5). */\nconst DEFAULT_MODEL_URL =\n  'https://cdn.jsdelivr.net/npm/@ricky0123/vad-web@0.0.20/dist/silero_vad_v5.onnx';\n\n/** Cache API key used for persisting the downloaded ONNX model. */\nconst CACHE_NAME = `guidekit-vad-v${VAD_VERSION}`;\nconst CACHE_MODEL_KEY = 'model.onnx';\n\n/** Silero VAD frame size: 512 samples at 16 kHz = 32 ms per frame. */\nconst FRAME_SIZE = 512;\n\n/** Target sample rate for VAD processing. */\nconst TARGET_SAMPLE_RATE = 16000;\n\n/** Duration (in ms) of audio collected for noise floor calibration. */\nconst CALIBRATION_DURATION_MS = 500;\n\n/** Hidden/cell state size for Silero VAD v5 LSTM. */\nconst STATE_SIZE = 128;\n\n// ---------------------------------------------------------------------------\n// Types\n// ---------------------------------------------------------------------------\n\nexport interface VADOptions {\n  /** Speech probability threshold (0-1). Default: 0.5 */\n  threshold?: number;\n  /** Minimum speech duration in ms to trigger start. Default: 300 */\n  minSpeechDurationMs?: number;\n  /** Silence duration in ms after speech to trigger end. Default: 500 */\n  silenceDurationMs?: number;\n  /** Sample rate. Default: 16000 */\n  sampleRate?: number;\n  /** Enable debug logging. Default: false */\n  debug?: boolean;\n  /** Custom URL for the Silero VAD ONNX model file. */\n  modelUrl?: string;\n}\n\nexport interface VADEvent {\n  type: 'speech-start' | 'speech-end' | 'vad-ready';\n  timestamp: number;\n  /** Speech probability (0-1) at the moment of the event. */\n  probability?: number;\n}\n\ntype VADEventType = VADEvent['type'];\ntype VADCallback = (event: VADEvent) => void;\n\n// ---------------------------------------------------------------------------\n// Utility: Cache API helpers\n// ---------------------------------------------------------------------------\n\nasync function loadModelFromCache(): Promise<ArrayBuffer | null> {\n  if (typeof caches === 'undefined') return null;\n  try {\n    const cache = await caches.open(CACHE_NAME);\n    const response = await cache.match(CACHE_MODEL_KEY);\n    return response ? response.arrayBuffer() : null;\n  } catch {\n    // Cache API may be unavailable in certain contexts (e.g. opaque origins).\n    return null;\n  }\n}\n\nasync function saveModelToCache(data: ArrayBuffer): Promise<void> {\n  if (typeof caches === 'undefined') return;\n  try {\n    const cache = await caches.open(CACHE_NAME);\n    await cache.put(CACHE_MODEL_KEY, new Response(data));\n  } catch {\n    // Silently ignore cache write failures.\n  }\n}\n\n// ---------------------------------------------------------------------------\n// Utility: Resampler\n// ---------------------------------------------------------------------------\n\n/**\n * Simple linear-interpolation resampler from `inputRate` to `outputRate`.\n * Adequate for VAD where perceptual audio quality is irrelevant.\n */\nfunction resample(\n  input: Float32Array,\n  inputRate: number,\n  outputRate: number,\n): Float32Array {\n  if (inputRate === outputRate) return input;\n  const ratio = inputRate / outputRate;\n  const outputLength = Math.round(input.length / ratio);\n  const output = new Float32Array(outputLength);\n  for (let i = 0; i < outputLength; i++) {\n    const srcIndex = i * ratio;\n    const srcFloor = Math.floor(srcIndex);\n    const srcCeil = Math.min(srcFloor + 1, input.length - 1);\n    const frac = srcIndex - srcFloor;\n    output[i] = (input[srcFloor] as number) * (1 - frac) + (input[srcCeil] as number) * frac;\n  }\n  return output;\n}\n\n// ---------------------------------------------------------------------------\n// SileroVAD\n// ---------------------------------------------------------------------------\n\nexport class SileroVAD {\n  // Options (resolved with defaults)\n  private readonly _threshold: number;\n  private readonly _minSpeechDurationMs: number;\n  private readonly _silenceDurationMs: number;\n  private readonly _sampleRate: number;\n  private readonly _debug: boolean;\n  private readonly _modelUrl: string;\n\n  // ONNX Runtime state\n  private _session: ort.InferenceSession | null = null;\n  private _h: ort.Tensor | null = null;\n  private _c: ort.Tensor | null = null;\n\n  // Audio pipeline\n  private _audioContext: AudioContext | null = null;\n  private _ownsAudioContext = false;\n  private _sourceNode: MediaStreamAudioSourceNode | null = null;\n  private _workletNode: AudioWorkletNode | ScriptProcessorNode | null = null;\n  private _stream: MediaStream | null = null;\n\n  // Frame buffer for accumulating resampled samples into FRAME_SIZE chunks\n  private _frameBuffer: Float32Array = new Float32Array(0);\n  private _frameBufferOffset = 0;\n\n  // State tracking\n  private _isReady = false;\n  private _isSpeaking = false;\n  private _isStarted = false;\n  private _isDestroyed = false;\n\n  // Duration tracking (in frames)\n  private _consecutiveSpeechFrames = 0;\n  private _consecutiveSilenceFrames = 0;\n  private _frameDurationMs: number;\n  private _minSpeechFrames: number;\n  private _silenceFrames: number;\n\n  // Noise floor calibration\n  private _isCalibrating = false;\n  private _calibrationSamples: number[] = [];\n  private _calibrationFramesNeeded = 0;\n  private _calibratedThreshold: number;\n\n  // Event listeners\n  private _listeners: Map<VADEventType, Set<VADCallback>> = new Map();\n\n  // Processing lock to serialise frame inference\n  private _processingPromise: Promise<void> = Promise.resolve();\n\n  constructor(options?: VADOptions) {\n    this._threshold = options?.threshold ?? 0.5;\n    this._minSpeechDurationMs = options?.minSpeechDurationMs ?? 300;\n    this._silenceDurationMs = options?.silenceDurationMs ?? 500;\n    this._sampleRate = options?.sampleRate ?? TARGET_SAMPLE_RATE;\n    this._debug = options?.debug ?? false;\n    this._modelUrl = options?.modelUrl ?? DEFAULT_MODEL_URL;\n    this._calibratedThreshold = this._threshold;\n\n    // Pre-compute frame-duration-based counters\n    this._frameDurationMs = (FRAME_SIZE / this._sampleRate) * 1000;\n    this._minSpeechFrames = Math.ceil(this._minSpeechDurationMs / this._frameDurationMs);\n    this._silenceFrames = Math.ceil(this._silenceDurationMs / this._frameDurationMs);\n\n    this._log('Created with options', {\n      threshold: this._threshold,\n      minSpeechDurationMs: this._minSpeechDurationMs,\n      silenceDurationMs: this._silenceDurationMs,\n      sampleRate: this._sampleRate,\n      modelUrl: this._modelUrl,\n    });\n  }\n\n  // -------------------------------------------------------------------------\n  // Public API\n  // -------------------------------------------------------------------------\n\n  /** Load the ONNX model. Uses Cache API for persistence across sessions. */\n  async init(): Promise<void> {\n    if (this._isDestroyed) {\n      throw new Error(`${LOG_PREFIX} Cannot init after destroy`);\n    }\n    if (this._isReady) {\n      this._log('Already initialised — skipping');\n      return;\n    }\n\n    this._log('Initialising...');\n\n    // 1. Attempt to load model bytes from cache, falling back to network.\n    let modelBuffer = await loadModelFromCache();\n    if (modelBuffer) {\n      this._log('Loaded model from Cache API');\n    } else {\n      this._log('Fetching model from', this._modelUrl);\n      const response = await fetch(this._modelUrl);\n      if (!response.ok) {\n        throw new Error(\n          `${LOG_PREFIX} Failed to fetch model: ${response.status} ${response.statusText}`,\n        );\n      }\n      modelBuffer = await response.arrayBuffer();\n      this._log('Model fetched, size:', modelBuffer.byteLength, 'bytes');\n\n      // Persist to Cache API for next time.\n      await saveModelToCache(modelBuffer);\n      this._log('Model saved to Cache API');\n    }\n\n    // 2. Create ONNX InferenceSession.\n    this._session = await ort.InferenceSession.create(modelBuffer, {\n      executionProviders: ['wasm'],\n      graphOptimizationLevel: 'all',\n    });\n\n    // 3. Initialise LSTM hidden/cell state tensors (zeros).\n    this._resetStates();\n\n    this._isReady = true;\n    this._log('Model loaded and ready');\n\n    this._emit({\n      type: 'vad-ready',\n      timestamp: Date.now(),\n    });\n  }\n\n  /**\n   * Process a single audio frame (512 samples at 16 kHz).\n   * Returns the speech probability (0-1).\n   */\n  async processFrame(audioData: Float32Array): Promise<number> {\n    if (!this._session) {\n      throw new Error(`${LOG_PREFIX} Model not loaded. Call init() first.`);\n    }\n    if (audioData.length !== FRAME_SIZE) {\n      throw new Error(\n        `${LOG_PREFIX} Expected ${FRAME_SIZE} samples, got ${audioData.length}`,\n      );\n    }\n\n    const inputTensor = new ort.Tensor('float32', audioData, [1, FRAME_SIZE]);\n    const srTensor = new ort.Tensor('int64', BigInt64Array.from([BigInt(this._sampleRate)]), [1]);\n\n    const feeds: Record<string, ort.Tensor> = {\n      input: inputTensor,\n      sr: srTensor,\n      h: this._h!,\n      c: this._c!,\n    };\n\n    const results = await this._session.run(feeds);\n\n    // Update LSTM hidden/cell states for the next frame.\n    this._h = results['hn'] as ort.Tensor;\n    this._c = results['cn'] as ort.Tensor;\n\n    const probability = (results['output'] as ort.Tensor).data[0] as number;\n    return probability;\n  }\n\n  /** Start VAD processing on a MediaStream (typically from getUserMedia). */\n  start(stream: MediaStream): void {\n    if (this._isDestroyed) {\n      throw new Error(`${LOG_PREFIX} Cannot start after destroy`);\n    }\n    if (!this._isReady) {\n      throw new Error(`${LOG_PREFIX} Model not loaded. Call init() first.`);\n    }\n    if (this._isStarted) {\n      this._log('Already started — stopping previous session first');\n      this.stop();\n    }\n\n    this._log('Starting VAD on MediaStream');\n    this._stream = stream;\n    this._isStarted = true;\n\n    // Reset speech tracking state.\n    this._isSpeaking = false;\n    this._consecutiveSpeechFrames = 0;\n    this._consecutiveSilenceFrames = 0;\n    this._frameBuffer = new Float32Array(FRAME_SIZE);\n    this._frameBufferOffset = 0;\n\n    // Reset LSTM states for a fresh stream.\n    this._resetStates();\n\n    // Begin noise floor calibration.\n    this._isCalibrating = true;\n    this._calibrationSamples = [];\n    this._calibrationFramesNeeded = Math.ceil(\n      (CALIBRATION_DURATION_MS / 1000) * this._sampleRate / FRAME_SIZE,\n    );\n    this._log('Calibrating noise floor for', this._calibrationFramesNeeded, 'frames');\n\n    // Build the audio processing pipeline.\n    this._setupAudioPipeline(stream);\n  }\n\n  /** Stop VAD processing and release audio resources (but keep the model). */\n  stop(): void {\n    if (!this._isStarted) return;\n\n    this._log('Stopping VAD');\n\n    // Tear down audio nodes.\n    this._teardownAudioPipeline();\n\n    // If we were speaking, emit speech-end.\n    if (this._isSpeaking) {\n      this._isSpeaking = false;\n      this._emit({\n        type: 'speech-end',\n        timestamp: Date.now(),\n        probability: 0,\n      });\n    }\n\n    // Reset state.\n    this._isStarted = false;\n    this._isSpeaking = false;\n    this._consecutiveSpeechFrames = 0;\n    this._consecutiveSilenceFrames = 0;\n    this._frameBufferOffset = 0;\n    this._isCalibrating = false;\n    this._calibrationSamples = [];\n    this._processingPromise = Promise.resolve();\n\n    this._log('VAD stopped');\n  }\n\n  /** Register a callback for speech-start events. Returns an unsubscribe function. */\n  onSpeechStart(callback: VADCallback): () => void {\n    return this._on('speech-start', callback);\n  }\n\n  /** Register a callback for speech-end events. Returns an unsubscribe function. */\n  onSpeechEnd(callback: VADCallback): () => void {\n    return this._on('speech-end', callback);\n  }\n\n  /** Register a callback for vad-ready events. Returns an unsubscribe function. */\n  onReady(callback: VADCallback): () => void {\n    return this._on('vad-ready', callback);\n  }\n\n  /** Whether the ONNX model is loaded and ready. */\n  get isReady(): boolean {\n    return this._isReady;\n  }\n\n  /** Whether speech is currently detected. */\n  get isSpeaking(): boolean {\n    return this._isSpeaking;\n  }\n\n  /** The MediaStream currently being processed, or null. */\n  get stream(): MediaStream | null {\n    return this._stream;\n  }\n\n  /** Release ONNX model session and all audio resources. */\n  async destroy(): Promise<void> {\n    if (this._isDestroyed) return;\n    this._log('Destroying...');\n\n    this.stop();\n\n    if (this._session) {\n      await this._session.release();\n      this._session = null;\n    }\n\n    // Dispose tensors.\n    this._h?.dispose();\n    this._c?.dispose();\n    this._h = null;\n    this._c = null;\n\n    this._isReady = false;\n    this._isDestroyed = true;\n    this._listeners.clear();\n\n    this._log('Destroyed');\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: Event system\n  // -------------------------------------------------------------------------\n\n  private _on(type: VADEventType, callback: VADCallback): () => void {\n    let set = this._listeners.get(type);\n    if (!set) {\n      set = new Set();\n      this._listeners.set(type, set);\n    }\n    set.add(callback);\n    return () => {\n      set!.delete(callback);\n    };\n  }\n\n  private _emit(event: VADEvent): void {\n    const set = this._listeners.get(event.type);\n    if (!set) return;\n    for (const cb of set) {\n      try {\n        cb(event);\n      } catch (err) {\n        // eslint-disable-next-line no-console\n        console.error(`${LOG_PREFIX} Error in ${event.type} callback:`, err);\n      }\n    }\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: Audio pipeline\n  // -------------------------------------------------------------------------\n\n  private _setupAudioPipeline(stream: MediaStream): void {\n    // Determine the incoming sample rate.\n    const tracks = stream.getAudioTracks();\n    const trackSettings = tracks[0]?.getSettings();\n    const inputSampleRate = trackSettings?.sampleRate ?? 48000;\n\n    this._log('Input sample rate:', inputSampleRate);\n\n    // Create AudioContext at the input sample rate so we don't double-resample.\n    // SSR guard: AudioContext may not exist.\n    if (typeof AudioContext === 'undefined' && typeof webkitAudioContext === 'undefined') {\n      throw new Error(`${LOG_PREFIX} AudioContext is not available in this environment`);\n    }\n\n    const AudioContextClass =\n      typeof AudioContext !== 'undefined'\n        ? AudioContext\n        : // eslint-disable-next-line @typescript-eslint/no-explicit-any\n          (globalThis as any).webkitAudioContext as typeof AudioContext;\n\n    this._audioContext = new AudioContextClass({ sampleRate: inputSampleRate });\n    this._ownsAudioContext = true;\n\n    this._sourceNode = this._audioContext.createMediaStreamSource(stream);\n\n    // Try AudioWorklet first, fall back to ScriptProcessorNode.\n    this._setupScriptProcessor(inputSampleRate);\n  }\n\n  /**\n   * ScriptProcessorNode fallback (works everywhere, including Safari).\n   * We use a buffer size of 4096 which gives ~85 ms of audio at 48 kHz.\n   */\n  private _setupScriptProcessor(inputSampleRate: number): void {\n    if (!this._audioContext || !this._sourceNode) return;\n\n    // Buffer size must be a power of 2: 256, 512, 1024, 2048, 4096, 8192, 16384.\n    const bufferSize = 4096;\n    const processor = this._audioContext.createScriptProcessor(bufferSize, 1, 1);\n\n    processor.onaudioprocess = (event: AudioProcessingEvent) => {\n      if (!this._isStarted) return;\n\n      const inputData = event.inputBuffer.getChannelData(0);\n\n      // Resample to target rate if needed.\n      const resampled =\n        inputSampleRate !== this._sampleRate\n          ? resample(inputData, inputSampleRate, this._sampleRate)\n          : new Float32Array(inputData);\n\n      // Feed resampled audio into frame-sized chunks.\n      this._feedAudio(resampled);\n    };\n\n    this._sourceNode.connect(processor);\n    processor.connect(this._audioContext.destination);\n    this._workletNode = processor;\n\n    this._log('Audio pipeline set up (ScriptProcessorNode)');\n  }\n\n  /**\n   * Accumulate resampled audio into FRAME_SIZE chunks and process each full frame.\n   */\n  private _feedAudio(samples: Float32Array): void {\n    let offset = 0;\n\n    while (offset < samples.length) {\n      const remaining = FRAME_SIZE - this._frameBufferOffset;\n      const available = samples.length - offset;\n      const toCopy = Math.min(remaining, available);\n\n      this._frameBuffer.set(\n        samples.subarray(offset, offset + toCopy),\n        this._frameBufferOffset,\n      );\n      this._frameBufferOffset += toCopy;\n      offset += toCopy;\n\n      if (this._frameBufferOffset === FRAME_SIZE) {\n        const frame = new Float32Array(this._frameBuffer);\n        this._frameBufferOffset = 0;\n\n        // Serialise inference calls to avoid overlapping ONNX sessions.\n        this._processingPromise = this._processingPromise.then(() =>\n          this._handleFrame(frame),\n        );\n      }\n    }\n  }\n\n  /**\n   * Process a single FRAME_SIZE frame: run inference and update speech state.\n   */\n  private async _handleFrame(frame: Float32Array): Promise<void> {\n    if (!this._isStarted || !this._session) return;\n\n    let probability: number;\n    try {\n      probability = await this.processFrame(frame);\n    } catch (err) {\n      if (this._debug) {\n        // eslint-disable-next-line no-console\n        console.error(`${LOG_PREFIX} Inference error:`, err);\n      }\n      return;\n    }\n\n    // Noise floor calibration phase.\n    if (this._isCalibrating) {\n      this._calibrationSamples.push(probability);\n\n      if (this._calibrationSamples.length >= this._calibrationFramesNeeded) {\n        this._finishCalibration();\n      }\n      return;\n    }\n\n    // Speech state machine.\n    const isSpeechFrame = probability >= this._calibratedThreshold;\n\n    if (isSpeechFrame) {\n      this._consecutiveSpeechFrames++;\n      this._consecutiveSilenceFrames = 0;\n\n      if (!this._isSpeaking && this._consecutiveSpeechFrames >= this._minSpeechFrames) {\n        this._isSpeaking = true;\n        this._log('Speech started, probability:', probability.toFixed(3));\n        this._emit({\n          type: 'speech-start',\n          timestamp: Date.now(),\n          probability,\n        });\n      }\n    } else {\n      this._consecutiveSilenceFrames++;\n      // Do NOT reset _consecutiveSpeechFrames here — only reset when speech-end fires.\n\n      if (this._isSpeaking && this._consecutiveSilenceFrames >= this._silenceFrames) {\n        this._isSpeaking = false;\n        this._consecutiveSpeechFrames = 0;\n        this._log('Speech ended, probability:', probability.toFixed(3));\n        this._emit({\n          type: 'speech-end',\n          timestamp: Date.now(),\n          probability,\n        });\n      }\n    }\n  }\n\n  private _finishCalibration(): void {\n    if (this._calibrationSamples.length === 0) {\n      this._isCalibrating = false;\n      return;\n    }\n\n    // Compute average noise floor probability.\n    const sum = this._calibrationSamples.reduce((a, b) => a + b, 0);\n    const avgNoise = sum / this._calibrationSamples.length;\n\n    // If the ambient noise floor is high, nudge the threshold above it.\n    // We add a margin so we don't constantly trigger on background noise.\n    const NOISE_MARGIN = 0.15;\n    if (avgNoise + NOISE_MARGIN > this._threshold) {\n      this._calibratedThreshold = Math.min(avgNoise + NOISE_MARGIN, 0.95);\n      this._log(\n        'Noise floor is high. Adjusted threshold from',\n        this._threshold.toFixed(3),\n        'to',\n        this._calibratedThreshold.toFixed(3),\n        '(avg noise:',\n        avgNoise.toFixed(3) + ')',\n      );\n    } else {\n      this._calibratedThreshold = this._threshold;\n      this._log('Noise floor OK, avg:', avgNoise.toFixed(3), '— keeping threshold at', this._threshold.toFixed(3));\n    }\n\n    // Recompute frame counters in case threshold changed min speech behaviour.\n    this._minSpeechFrames = Math.ceil(this._minSpeechDurationMs / this._frameDurationMs);\n    this._silenceFrames = Math.ceil(this._silenceDurationMs / this._frameDurationMs);\n\n    this._isCalibrating = false;\n    this._calibrationSamples = [];\n  }\n\n  private _teardownAudioPipeline(): void {\n    if (this._workletNode) {\n      try {\n        this._workletNode.disconnect();\n      } catch {\n        // Ignore disconnect errors.\n      }\n      if ('onaudioprocess' in this._workletNode) {\n        (this._workletNode as ScriptProcessorNode).onaudioprocess = null;\n      }\n      this._workletNode = null;\n    }\n\n    if (this._sourceNode) {\n      try {\n        this._sourceNode.disconnect();\n      } catch {\n        // Ignore.\n      }\n      this._sourceNode = null;\n    }\n\n    if (this._audioContext && this._ownsAudioContext) {\n      try {\n        void this._audioContext.close();\n      } catch {\n        // Ignore.\n      }\n      this._audioContext = null;\n      this._ownsAudioContext = false;\n    }\n\n    this._stream = null;\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: ONNX state helpers\n  // -------------------------------------------------------------------------\n\n  /** Reset the LSTM hidden and cell states to zeros. */\n  private _resetStates(): void {\n    // Dispose any existing tensors to free memory.\n    this._h?.dispose();\n    this._c?.dispose();\n\n    const zeros = new Float32Array(2 * STATE_SIZE).fill(0);\n    this._h = new ort.Tensor('float32', zeros.slice(0, STATE_SIZE), [2, 1, 64]);\n    this._c = new ort.Tensor('float32', zeros.slice(STATE_SIZE), [2, 1, 64]);\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: Logging\n  // -------------------------------------------------------------------------\n\n  private _log(...args: unknown[]): void {\n    if (!this._debug) return;\n    // eslint-disable-next-line no-console\n    console.log(LOG_PREFIX, ...args);\n  }\n}\n\n// ---------------------------------------------------------------------------\n// Convenience factory\n// ---------------------------------------------------------------------------\n\n/**\n * Create and initialise a SileroVAD instance in one call.\n *\n * ```ts\n * const vad = await createVAD({ debug: true });\n * vad.onSpeechStart(() => console.log('speaking'));\n * vad.start(stream);\n * ```\n */\nexport async function createVAD(options?: VADOptions): Promise<SileroVAD> {\n  const vad = new SileroVAD(options);\n  await vad.init();\n  return vad;\n}\n\n// Re-export the frame size constant so consumers can align their buffers.\nexport { FRAME_SIZE, TARGET_SAMPLE_RATE };\n\n// Type-only declaration for environments that provide webkitAudioContext.\ndeclare global {\n  // eslint-disable-next-line no-var\n  var webkitAudioContext: typeof AudioContext | undefined;\n}\n"],"mappings":";AACA,YAAY,SAAS;AAEd,IAAM,cAAc;AAM3B,IAAM,aAAa;AAGnB,IAAM,oBACJ;AAGF,IAAM,aAAa,iBAAiB,WAAW;AAC/C,IAAM,kBAAkB;AAGxB,IAAM,aAAa;AAGnB,IAAM,qBAAqB;AAG3B,IAAM,0BAA0B;AAGhC,IAAM,aAAa;AAmCnB,eAAe,qBAAkD;AAC/D,MAAI,OAAO,WAAW,YAAa,QAAO;AAC1C,MAAI;AACF,UAAM,QAAQ,MAAM,OAAO,KAAK,UAAU;AAC1C,UAAM,WAAW,MAAM,MAAM,MAAM,eAAe;AAClD,WAAO,WAAW,SAAS,YAAY,IAAI;AAAA,EAC7C,QAAQ;AAEN,WAAO;AAAA,EACT;AACF;AAEA,eAAe,iBAAiB,MAAkC;AAChE,MAAI,OAAO,WAAW,YAAa;AACnC,MAAI;AACF,UAAM,QAAQ,MAAM,OAAO,KAAK,UAAU;AAC1C,UAAM,MAAM,IAAI,iBAAiB,IAAI,SAAS,IAAI,CAAC;AAAA,EACrD,QAAQ;AAAA,EAER;AACF;AAUA,SAAS,SACP,OACA,WACA,YACc;AACd,MAAI,cAAc,WAAY,QAAO;AACrC,QAAM,QAAQ,YAAY;AAC1B,QAAM,eAAe,KAAK,MAAM,MAAM,SAAS,KAAK;AACpD,QAAM,SAAS,IAAI,aAAa,YAAY;AAC5C,WAAS,IAAI,GAAG,IAAI,cAAc,KAAK;AACrC,UAAM,WAAW,IAAI;AACrB,UAAM,WAAW,KAAK,MAAM,QAAQ;AACpC,UAAM,UAAU,KAAK,IAAI,WAAW,GAAG,MAAM,SAAS,CAAC;AACvD,UAAM,OAAO,WAAW;AACxB,WAAO,CAAC,IAAK,MAAM,QAAQ,KAAgB,IAAI,QAAS,MAAM,OAAO,IAAe;AAAA,EACtF;AACA,SAAO;AACT;AAMO,IAAM,YAAN,MAAgB;AAAA;AAAA,EAEJ;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA;AAAA,EAGT,WAAwC;AAAA,EACxC,KAAwB;AAAA,EACxB,KAAwB;AAAA;AAAA,EAGxB,gBAAqC;AAAA,EACrC,oBAAoB;AAAA,EACpB,cAAiD;AAAA,EACjD,eAA8D;AAAA,EAC9D,UAA8B;AAAA;AAAA,EAG9B,eAA6B,IAAI,aAAa,CAAC;AAAA,EAC/C,qBAAqB;AAAA;AAAA,EAGrB,WAAW;AAAA,EACX,cAAc;AAAA,EACd,aAAa;AAAA,EACb,eAAe;AAAA;AAAA,EAGf,2BAA2B;AAAA,EAC3B,4BAA4B;AAAA,EAC5B;AAAA,EACA;AAAA,EACA;AAAA;AAAA,EAGA,iBAAiB;AAAA,EACjB,sBAAgC,CAAC;AAAA,EACjC,2BAA2B;AAAA,EAC3B;AAAA;AAAA,EAGA,aAAkD,oBAAI,IAAI;AAAA;AAAA,EAG1D,qBAAoC,QAAQ,QAAQ;AAAA,EAE5D,YAAY,SAAsB;AAChC,SAAK,aAAa,SAAS,aAAa;AACxC,SAAK,uBAAuB,SAAS,uBAAuB;AAC5D,SAAK,qBAAqB,SAAS,qBAAqB;AACxD,SAAK,cAAc,SAAS,cAAc;AAC1C,SAAK,SAAS,SAAS,SAAS;AAChC,SAAK,YAAY,SAAS,YAAY;AACtC,SAAK,uBAAuB,KAAK;AAGjC,SAAK,mBAAoB,aAAa,KAAK,cAAe;AAC1D,SAAK,mBAAmB,KAAK,KAAK,KAAK,uBAAuB,KAAK,gBAAgB;AACnF,SAAK,iBAAiB,KAAK,KAAK,KAAK,qBAAqB,KAAK,gBAAgB;AAE/E,SAAK,KAAK,wBAAwB;AAAA,MAChC,WAAW,KAAK;AAAA,MAChB,qBAAqB,KAAK;AAAA,MAC1B,mBAAmB,KAAK;AAAA,MACxB,YAAY,KAAK;AAAA,MACjB,UAAU,KAAK;AAAA,IACjB,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,OAAsB;AAC1B,QAAI,KAAK,cAAc;AACrB,YAAM,IAAI,MAAM,GAAG,UAAU,4BAA4B;AAAA,IAC3D;AACA,QAAI,KAAK,UAAU;AACjB,WAAK,KAAK,qCAAgC;AAC1C;AAAA,IACF;AAEA,SAAK,KAAK,iBAAiB;AAG3B,QAAI,cAAc,MAAM,mBAAmB;AAC3C,QAAI,aAAa;AACf,WAAK,KAAK,6BAA6B;AAAA,IACzC,OAAO;AACL,WAAK,KAAK,uBAAuB,KAAK,SAAS;AAC/C,YAAM,WAAW,MAAM,MAAM,KAAK,SAAS;AAC3C,UAAI,CAAC,SAAS,IAAI;AAChB,cAAM,IAAI;AAAA,UACR,GAAG,UAAU,2BAA2B,SAAS,MAAM,IAAI,SAAS,UAAU;AAAA,QAChF;AAAA,MACF;AACA,oBAAc,MAAM,SAAS,YAAY;AACzC,WAAK,KAAK,wBAAwB,YAAY,YAAY,OAAO;AAGjE,YAAM,iBAAiB,WAAW;AAClC,WAAK,KAAK,0BAA0B;AAAA,IACtC;AAGA,SAAK,WAAW,MAAU,qBAAiB,OAAO,aAAa;AAAA,MAC7D,oBAAoB,CAAC,MAAM;AAAA,MAC3B,wBAAwB;AAAA,IAC1B,CAAC;AAGD,SAAK,aAAa;AAElB,SAAK,WAAW;AAChB,SAAK,KAAK,wBAAwB;AAElC,SAAK,MAAM;AAAA,MACT,MAAM;AAAA,MACN,WAAW,KAAK,IAAI;AAAA,IACtB,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,aAAa,WAA0C;AAC3D,QAAI,CAAC,KAAK,UAAU;AAClB,YAAM,IAAI,MAAM,GAAG,UAAU,uCAAuC;AAAA,IACtE;AACA,QAAI,UAAU,WAAW,YAAY;AACnC,YAAM,IAAI;AAAA,QACR,GAAG,UAAU,aAAa,UAAU,iBAAiB,UAAU,MAAM;AAAA,MACvE;AAAA,IACF;AAEA,UAAM,cAAc,IAAQ,WAAO,WAAW,WAAW,CAAC,GAAG,UAAU,CAAC;AACxE,UAAM,WAAW,IAAQ,WAAO,SAAS,cAAc,KAAK,CAAC,OAAO,KAAK,WAAW,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;AAE5F,UAAM,QAAoC;AAAA,MACxC,OAAO;AAAA,MACP,IAAI;AAAA,MACJ,GAAG,KAAK;AAAA,MACR,GAAG,KAAK;AAAA,IACV;AAEA,UAAM,UAAU,MAAM,KAAK,SAAS,IAAI,KAAK;AAG7C,SAAK,KAAK,QAAQ,IAAI;AACtB,SAAK,KAAK,QAAQ,IAAI;AAEtB,UAAM,cAAe,QAAQ,QAAQ,EAAiB,KAAK,CAAC;AAC5D,WAAO;AAAA,EACT;AAAA;AAAA,EAGA,MAAM,QAA2B;AAC/B,QAAI,KAAK,cAAc;AACrB,YAAM,IAAI,MAAM,GAAG,UAAU,6BAA6B;AAAA,IAC5D;AACA,QAAI,CAAC,KAAK,UAAU;AAClB,YAAM,IAAI,MAAM,GAAG,UAAU,uCAAuC;AAAA,IACtE;AACA,QAAI,KAAK,YAAY;AACnB,WAAK,KAAK,wDAAmD;AAC7D,WAAK,KAAK;AAAA,IACZ;AAEA,SAAK,KAAK,6BAA6B;AACvC,SAAK,UAAU;AACf,SAAK,aAAa;AAGlB,SAAK,cAAc;AACnB,SAAK,2BAA2B;AAChC,SAAK,4BAA4B;AACjC,SAAK,eAAe,IAAI,aAAa,UAAU;AAC/C,SAAK,qBAAqB;AAG1B,SAAK,aAAa;AAGlB,SAAK,iBAAiB;AACtB,SAAK,sBAAsB,CAAC;AAC5B,SAAK,2BAA2B,KAAK;AAAA,MAClC,0BAA0B,MAAQ,KAAK,cAAc;AAAA,IACxD;AACA,SAAK,KAAK,+BAA+B,KAAK,0BAA0B,QAAQ;AAGhF,SAAK,oBAAoB,MAAM;AAAA,EACjC;AAAA;AAAA,EAGA,OAAa;AACX,QAAI,CAAC,KAAK,WAAY;AAEtB,SAAK,KAAK,cAAc;AAGxB,SAAK,uBAAuB;AAG5B,QAAI,KAAK,aAAa;AACpB,WAAK,cAAc;AACnB,WAAK,MAAM;AAAA,QACT,MAAM;AAAA,QACN,WAAW,KAAK,IAAI;AAAA,QACpB,aAAa;AAAA,MACf,CAAC;AAAA,IACH;AAGA,SAAK,aAAa;AAClB,SAAK,cAAc;AACnB,SAAK,2BAA2B;AAChC,SAAK,4BAA4B;AACjC,SAAK,qBAAqB;AAC1B,SAAK,iBAAiB;AACtB,SAAK,sBAAsB,CAAC;AAC5B,SAAK,qBAAqB,QAAQ,QAAQ;AAE1C,SAAK,KAAK,aAAa;AAAA,EACzB;AAAA;AAAA,EAGA,cAAc,UAAmC;AAC/C,WAAO,KAAK,IAAI,gBAAgB,QAAQ;AAAA,EAC1C;AAAA;AAAA,EAGA,YAAY,UAAmC;AAC7C,WAAO,KAAK,IAAI,cAAc,QAAQ;AAAA,EACxC;AAAA;AAAA,EAGA,QAAQ,UAAmC;AACzC,WAAO,KAAK,IAAI,aAAa,QAAQ;AAAA,EACvC;AAAA;AAAA,EAGA,IAAI,UAAmB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,aAAsB;AACxB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,SAA6B;AAC/B,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,MAAM,UAAyB;AAC7B,QAAI,KAAK,aAAc;AACvB,SAAK,KAAK,eAAe;AAEzB,SAAK,KAAK;AAEV,QAAI,KAAK,UAAU;AACjB,YAAM,KAAK,SAAS,QAAQ;AAC5B,WAAK,WAAW;AAAA,IAClB;AAGA,SAAK,IAAI,QAAQ;AACjB,SAAK,IAAI,QAAQ;AACjB,SAAK,KAAK;AACV,SAAK,KAAK;AAEV,SAAK,WAAW;AAChB,SAAK,eAAe;AACpB,SAAK,WAAW,MAAM;AAEtB,SAAK,KAAK,WAAW;AAAA,EACvB;AAAA;AAAA;AAAA;AAAA,EAMQ,IAAI,MAAoB,UAAmC;AACjE,QAAI,MAAM,KAAK,WAAW,IAAI,IAAI;AAClC,QAAI,CAAC,KAAK;AACR,YAAM,oBAAI,IAAI;AACd,WAAK,WAAW,IAAI,MAAM,GAAG;AAAA,IAC/B;AACA,QAAI,IAAI,QAAQ;AAChB,WAAO,MAAM;AACX,UAAK,OAAO,QAAQ;AAAA,IACtB;AAAA,EACF;AAAA,EAEQ,MAAM,OAAuB;AACnC,UAAM,MAAM,KAAK,WAAW,IAAI,MAAM,IAAI;AAC1C,QAAI,CAAC,IAAK;AACV,eAAW,MAAM,KAAK;AACpB,UAAI;AACF,WAAG,KAAK;AAAA,MACV,SAAS,KAAK;AAEZ,gBAAQ,MAAM,GAAG,UAAU,aAAa,MAAM,IAAI,cAAc,GAAG;AAAA,MACrE;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAMQ,oBAAoB,QAA2B;AAErD,UAAM,SAAS,OAAO,eAAe;AACrC,UAAM,gBAAgB,OAAO,CAAC,GAAG,YAAY;AAC7C,UAAM,kBAAkB,eAAe,cAAc;AAErD,SAAK,KAAK,sBAAsB,eAAe;AAI/C,QAAI,OAAO,iBAAiB,eAAe,OAAO,uBAAuB,aAAa;AACpF,YAAM,IAAI,MAAM,GAAG,UAAU,oDAAoD;AAAA,IACnF;AAEA,UAAM,oBACJ,OAAO,iBAAiB,cACpB;AAAA;AAAA,MAEC,WAAmB;AAAA;AAE1B,SAAK,gBAAgB,IAAI,kBAAkB,EAAE,YAAY,gBAAgB,CAAC;AAC1E,SAAK,oBAAoB;AAEzB,SAAK,cAAc,KAAK,cAAc,wBAAwB,MAAM;AAGpE,SAAK,sBAAsB,eAAe;AAAA,EAC5C;AAAA;AAAA;AAAA;AAAA;AAAA,EAMQ,sBAAsB,iBAA+B;AAC3D,QAAI,CAAC,KAAK,iBAAiB,CAAC,KAAK,YAAa;AAG9C,UAAM,aAAa;AACnB,UAAM,YAAY,KAAK,cAAc,sBAAsB,YAAY,GAAG,CAAC;AAE3E,cAAU,iBAAiB,CAAC,UAAgC;AAC1D,UAAI,CAAC,KAAK,WAAY;AAEtB,YAAM,YAAY,MAAM,YAAY,eAAe,CAAC;AAGpD,YAAM,YACJ,oBAAoB,KAAK,cACrB,SAAS,WAAW,iBAAiB,KAAK,WAAW,IACrD,IAAI,aAAa,SAAS;AAGhC,WAAK,WAAW,SAAS;AAAA,IAC3B;AAEA,SAAK,YAAY,QAAQ,SAAS;AAClC,cAAU,QAAQ,KAAK,cAAc,WAAW;AAChD,SAAK,eAAe;AAEpB,SAAK,KAAK,6CAA6C;AAAA,EACzD;AAAA;AAAA;AAAA;AAAA,EAKQ,WAAW,SAA6B;AAC9C,QAAI,SAAS;AAEb,WAAO,SAAS,QAAQ,QAAQ;AAC9B,YAAM,YAAY,aAAa,KAAK;AACpC,YAAM,YAAY,QAAQ,SAAS;AACnC,YAAM,SAAS,KAAK,IAAI,WAAW,SAAS;AAE5C,WAAK,aAAa;AAAA,QAChB,QAAQ,SAAS,QAAQ,SAAS,MAAM;AAAA,QACxC,KAAK;AAAA,MACP;AACA,WAAK,sBAAsB;AAC3B,gBAAU;AAEV,UAAI,KAAK,uBAAuB,YAAY;AAC1C,cAAM,QAAQ,IAAI,aAAa,KAAK,YAAY;AAChD,aAAK,qBAAqB;AAG1B,aAAK,qBAAqB,KAAK,mBAAmB;AAAA,UAAK,MACrD,KAAK,aAAa,KAAK;AAAA,QACzB;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,MAAc,aAAa,OAAoC;AAC7D,QAAI,CAAC,KAAK,cAAc,CAAC,KAAK,SAAU;AAExC,QAAI;AACJ,QAAI;AACF,oBAAc,MAAM,KAAK,aAAa,KAAK;AAAA,IAC7C,SAAS,KAAK;AACZ,UAAI,KAAK,QAAQ;AAEf,gBAAQ,MAAM,GAAG,UAAU,qBAAqB,GAAG;AAAA,MACrD;AACA;AAAA,IACF;AAGA,QAAI,KAAK,gBAAgB;AACvB,WAAK,oBAAoB,KAAK,WAAW;AAEzC,UAAI,KAAK,oBAAoB,UAAU,KAAK,0BAA0B;AACpE,aAAK,mBAAmB;AAAA,MAC1B;AACA;AAAA,IACF;AAGA,UAAM,gBAAgB,eAAe,KAAK;AAE1C,QAAI,eAAe;AACjB,WAAK;AACL,WAAK,4BAA4B;AAEjC,UAAI,CAAC,KAAK,eAAe,KAAK,4BAA4B,KAAK,kBAAkB;AAC/E,aAAK,cAAc;AACnB,aAAK,KAAK,gCAAgC,YAAY,QAAQ,CAAC,CAAC;AAChE,aAAK,MAAM;AAAA,UACT,MAAM;AAAA,UACN,WAAW,KAAK,IAAI;AAAA,UACpB;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF,OAAO;AACL,WAAK;AAGL,UAAI,KAAK,eAAe,KAAK,6BAA6B,KAAK,gBAAgB;AAC7E,aAAK,cAAc;AACnB,aAAK,2BAA2B;AAChC,aAAK,KAAK,8BAA8B,YAAY,QAAQ,CAAC,CAAC;AAC9D,aAAK,MAAM;AAAA,UACT,MAAM;AAAA,UACN,WAAW,KAAK,IAAI;AAAA,UACpB;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,qBAA2B;AACjC,QAAI,KAAK,oBAAoB,WAAW,GAAG;AACzC,WAAK,iBAAiB;AACtB;AAAA,IACF;AAGA,UAAM,MAAM,KAAK,oBAAoB,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC9D,UAAM,WAAW,MAAM,KAAK,oBAAoB;AAIhD,UAAM,eAAe;AACrB,QAAI,WAAW,eAAe,KAAK,YAAY;AAC7C,WAAK,uBAAuB,KAAK,IAAI,WAAW,cAAc,IAAI;AAClE,WAAK;AAAA,QACH;AAAA,QACA,KAAK,WAAW,QAAQ,CAAC;AAAA,QACzB;AAAA,QACA,KAAK,qBAAqB,QAAQ,CAAC;AAAA,QACnC;AAAA,QACA,SAAS,QAAQ,CAAC,IAAI;AAAA,MACxB;AAAA,IACF,OAAO;AACL,WAAK,uBAAuB,KAAK;AACjC,WAAK,KAAK,wBAAwB,SAAS,QAAQ,CAAC,GAAG,+BAA0B,KAAK,WAAW,QAAQ,CAAC,CAAC;AAAA,IAC7G;AAGA,SAAK,mBAAmB,KAAK,KAAK,KAAK,uBAAuB,KAAK,gBAAgB;AACnF,SAAK,iBAAiB,KAAK,KAAK,KAAK,qBAAqB,KAAK,gBAAgB;AAE/E,SAAK,iBAAiB;AACtB,SAAK,sBAAsB,CAAC;AAAA,EAC9B;AAAA,EAEQ,yBAA+B;AACrC,QAAI,KAAK,cAAc;AACrB,UAAI;AACF,aAAK,aAAa,WAAW;AAAA,MAC/B,QAAQ;AAAA,MAER;AACA,UAAI,oBAAoB,KAAK,cAAc;AACzC,QAAC,KAAK,aAAqC,iBAAiB;AAAA,MAC9D;AACA,WAAK,eAAe;AAAA,IACtB;AAEA,QAAI,KAAK,aAAa;AACpB,UAAI;AACF,aAAK,YAAY,WAAW;AAAA,MAC9B,QAAQ;AAAA,MAER;AACA,WAAK,cAAc;AAAA,IACrB;AAEA,QAAI,KAAK,iBAAiB,KAAK,mBAAmB;AAChD,UAAI;AACF,aAAK,KAAK,cAAc,MAAM;AAAA,MAChC,QAAQ;AAAA,MAER;AACA,WAAK,gBAAgB;AACrB,WAAK,oBAAoB;AAAA,IAC3B;AAEA,SAAK,UAAU;AAAA,EACjB;AAAA;AAAA;AAAA;AAAA;AAAA,EAOQ,eAAqB;AAE3B,SAAK,IAAI,QAAQ;AACjB,SAAK,IAAI,QAAQ;AAEjB,UAAM,QAAQ,IAAI,aAAa,IAAI,UAAU,EAAE,KAAK,CAAC;AACrD,SAAK,KAAK,IAAQ,WAAO,WAAW,MAAM,MAAM,GAAG,UAAU,GAAG,CAAC,GAAG,GAAG,EAAE,CAAC;AAC1E,SAAK,KAAK,IAAQ,WAAO,WAAW,MAAM,MAAM,UAAU,GAAG,CAAC,GAAG,GAAG,EAAE,CAAC;AAAA,EACzE;AAAA;AAAA;AAAA;AAAA,EAMQ,QAAQ,MAAuB;AACrC,QAAI,CAAC,KAAK,OAAQ;AAElB,YAAQ,IAAI,YAAY,GAAG,IAAI;AAAA,EACjC;AACF;AAeA,eAAsB,UAAU,SAA0C;AACxE,QAAM,MAAM,IAAI,UAAU,OAAO;AACjC,QAAM,IAAI,KAAK;AACf,SAAO;AACT;","names":[]}
+{"version":3,"sources":["../src/index.ts"],"sourcesContent":["// @guidekit/vad — Silero VAD ONNX model wrapper for voice activity detection\nimport * as ort from 'onnxruntime-web';\n\nexport const VAD_VERSION = '0.1.0-beta.2';\n\n// ---------------------------------------------------------------------------\n// Constants\n// ---------------------------------------------------------------------------\n\nconst LOG_PREFIX = '[GuideKit:VAD]';\n\n/** Default CDN URL for the Silero VAD ONNX model (v5). */\nconst DEFAULT_MODEL_URL =\n  'https://cdn.jsdelivr.net/npm/@ricky0123/vad-web@0.0.20/dist/silero_vad_v5.onnx';\n\n/** Cache API key used for persisting the downloaded ONNX model. */\nconst CACHE_NAME = `guidekit-vad-v${VAD_VERSION}`;\nconst CACHE_MODEL_KEY = 'model.onnx';\n\n/** Silero VAD frame size: 512 samples at 16 kHz = 32 ms per frame. */\nconst FRAME_SIZE = 512;\n\n/** Target sample rate for VAD processing. */\nconst TARGET_SAMPLE_RATE = 16000;\n\n/** Duration (in ms) of audio collected for noise floor calibration. */\nconst CALIBRATION_DURATION_MS = 500;\n\n/** Hidden/cell state size for Silero VAD v5 LSTM. */\nconst STATE_SIZE = 128;\n\n// ---------------------------------------------------------------------------\n// Types\n// ---------------------------------------------------------------------------\n\nexport interface VADOptions {\n  /** Speech probability threshold (0-1). Default: 0.5 */\n  threshold?: number;\n  /** Minimum speech duration in ms to trigger start. Default: 300 */\n  minSpeechDurationMs?: number;\n  /** Silence duration in ms after speech to trigger end. Default: 500 */\n  silenceDurationMs?: number;\n  /** Sample rate. Default: 16000 */\n  sampleRate?: number;\n  /** Enable debug logging. Default: false */\n  debug?: boolean;\n  /** Custom URL for the Silero VAD ONNX model file. */\n  modelUrl?: string;\n}\n\nexport interface VADEvent {\n  type: 'speech-start' | 'speech-end' | 'vad-ready';\n  timestamp: number;\n  /** Speech probability (0-1) at the moment of the event. */\n  probability?: number;\n}\n\ntype VADEventType = VADEvent['type'];\ntype VADCallback = (event: VADEvent) => void;\n\n// ---------------------------------------------------------------------------\n// Utility: Cache API helpers\n// ---------------------------------------------------------------------------\n\nasync function loadModelFromCache(): Promise<ArrayBuffer | null> {\n  if (typeof caches === 'undefined') return null;\n  try {\n    const cache = await caches.open(CACHE_NAME);\n    const response = await cache.match(CACHE_MODEL_KEY);\n    return response ? response.arrayBuffer() : null;\n  } catch {\n    // Cache API may be unavailable in certain contexts (e.g. opaque origins).\n    return null;\n  }\n}\n\nasync function saveModelToCache(data: ArrayBuffer): Promise<void> {\n  if (typeof caches === 'undefined') return;\n  try {\n    const cache = await caches.open(CACHE_NAME);\n    await cache.put(CACHE_MODEL_KEY, new Response(data));\n  } catch {\n    // Silently ignore cache write failures.\n  }\n}\n\n// ---------------------------------------------------------------------------\n// Utility: Resampler\n// ---------------------------------------------------------------------------\n\n/**\n * Simple linear-interpolation resampler from `inputRate` to `outputRate`.\n * Adequate for VAD where perceptual audio quality is irrelevant.\n */\nfunction resample(\n  input: Float32Array,\n  inputRate: number,\n  outputRate: number,\n): Float32Array {\n  if (inputRate === outputRate) return input;\n  const ratio = inputRate / outputRate;\n  const outputLength = Math.round(input.length / ratio);\n  const output = new Float32Array(outputLength);\n  for (let i = 0; i < outputLength; i++) {\n    const srcIndex = i * ratio;\n    const srcFloor = Math.floor(srcIndex);\n    const srcCeil = Math.min(srcFloor + 1, input.length - 1);\n    const frac = srcIndex - srcFloor;\n    output[i] = (input[srcFloor] as number) * (1 - frac) + (input[srcCeil] as number) * frac;\n  }\n  return output;\n}\n\n// ---------------------------------------------------------------------------\n// SileroVAD\n// ---------------------------------------------------------------------------\n\nexport class SileroVAD {\n  // Options (resolved with defaults)\n  private readonly _threshold: number;\n  private readonly _minSpeechDurationMs: number;\n  private readonly _silenceDurationMs: number;\n  private readonly _sampleRate: number;\n  private readonly _debug: boolean;\n  private readonly _modelUrl: string;\n\n  // ONNX Runtime state\n  private _session: ort.InferenceSession | null = null;\n  private _h: ort.Tensor | null = null;\n  private _c: ort.Tensor | null = null;\n\n  // Audio pipeline\n  private _audioContext: AudioContext | null = null;\n  private _ownsAudioContext = false;\n  private _sourceNode: MediaStreamAudioSourceNode | null = null;\n  private _workletNode: AudioWorkletNode | ScriptProcessorNode | null = null;\n  private _stream: MediaStream | null = null;\n\n  // Frame buffer for accumulating resampled samples into FRAME_SIZE chunks\n  private _frameBuffer: Float32Array = new Float32Array(0);\n  private _frameBufferOffset = 0;\n\n  // State tracking\n  private _isReady = false;\n  private _isSpeaking = false;\n  private _isStarted = false;\n  private _isDestroyed = false;\n\n  // Duration tracking (in frames)\n  private _consecutiveSpeechFrames = 0;\n  private _consecutiveSilenceFrames = 0;\n  private _frameDurationMs: number;\n  private _minSpeechFrames: number;\n  private _silenceFrames: number;\n\n  // Noise floor calibration\n  private _isCalibrating = false;\n  private _calibrationSamples: number[] = [];\n  private _calibrationFramesNeeded = 0;\n  private _calibratedThreshold: number;\n\n  // Event listeners\n  private _listeners: Map<VADEventType, Set<VADCallback>> = new Map();\n\n  // Processing lock to serialise frame inference\n  private _processingPromise: Promise<void> = Promise.resolve();\n\n  constructor(options?: VADOptions) {\n    this._threshold = options?.threshold ?? 0.5;\n    this._minSpeechDurationMs = options?.minSpeechDurationMs ?? 300;\n    this._silenceDurationMs = options?.silenceDurationMs ?? 500;\n    this._sampleRate = options?.sampleRate ?? TARGET_SAMPLE_RATE;\n    this._debug = options?.debug ?? false;\n    this._modelUrl = options?.modelUrl ?? DEFAULT_MODEL_URL;\n    this._calibratedThreshold = this._threshold;\n\n    // Pre-compute frame-duration-based counters\n    this._frameDurationMs = (FRAME_SIZE / this._sampleRate) * 1000;\n    this._minSpeechFrames = Math.ceil(this._minSpeechDurationMs / this._frameDurationMs);\n    this._silenceFrames = Math.ceil(this._silenceDurationMs / this._frameDurationMs);\n\n    this._log('Created with options', {\n      threshold: this._threshold,\n      minSpeechDurationMs: this._minSpeechDurationMs,\n      silenceDurationMs: this._silenceDurationMs,\n      sampleRate: this._sampleRate,\n      modelUrl: this._modelUrl,\n    });\n  }\n\n  // -------------------------------------------------------------------------\n  // Public API\n  // -------------------------------------------------------------------------\n\n  /** Load the ONNX model. Uses Cache API for persistence across sessions. */\n  async init(): Promise<void> {\n    if (this._isDestroyed) {\n      throw new Error(`${LOG_PREFIX} Cannot init after destroy`);\n    }\n    if (this._isReady) {\n      this._log('Already initialised — skipping');\n      return;\n    }\n\n    this._log('Initialising...');\n\n    // 1. Attempt to load model bytes from cache, falling back to network.\n    let modelBuffer = await loadModelFromCache();\n    if (modelBuffer) {\n      this._log('Loaded model from Cache API');\n    } else {\n      this._log('Fetching model from', this._modelUrl);\n      const response = await fetch(this._modelUrl);\n      if (!response.ok) {\n        throw new Error(\n          `${LOG_PREFIX} Failed to fetch model: ${response.status} ${response.statusText}`,\n        );\n      }\n      modelBuffer = await response.arrayBuffer();\n      this._log('Model fetched, size:', modelBuffer.byteLength, 'bytes');\n\n      // Persist to Cache API for next time.\n      await saveModelToCache(modelBuffer);\n      this._log('Model saved to Cache API');\n    }\n\n    // 2. Create ONNX InferenceSession.\n    this._session = await ort.InferenceSession.create(modelBuffer, {\n      executionProviders: ['wasm'],\n      graphOptimizationLevel: 'all',\n    });\n\n    // 3. Initialise LSTM hidden/cell state tensors (zeros).\n    this._resetStates();\n\n    this._isReady = true;\n    this._log('Model loaded and ready');\n\n    this._emit({\n      type: 'vad-ready',\n      timestamp: Date.now(),\n    });\n  }\n\n  /**\n   * Process a single audio frame (512 samples at 16 kHz).\n   * Returns the speech probability (0-1).\n   */\n  async processFrame(audioData: Float32Array): Promise<number> {\n    if (!this._session) {\n      throw new Error(`${LOG_PREFIX} Model not loaded. Call init() first.`);\n    }\n    if (audioData.length !== FRAME_SIZE) {\n      throw new Error(\n        `${LOG_PREFIX} Expected ${FRAME_SIZE} samples, got ${audioData.length}`,\n      );\n    }\n\n    const inputTensor = new ort.Tensor('float32', audioData, [1, FRAME_SIZE]);\n    const srTensor = new ort.Tensor('int64', BigInt64Array.from([BigInt(this._sampleRate)]), [1]);\n\n    const feeds: Record<string, ort.Tensor> = {\n      input: inputTensor,\n      sr: srTensor,\n      h: this._h!,\n      c: this._c!,\n    };\n\n    const results = await this._session.run(feeds);\n\n    // Update LSTM hidden/cell states for the next frame.\n    this._h = results['hn'] as ort.Tensor;\n    this._c = results['cn'] as ort.Tensor;\n\n    const probability = (results['output'] as ort.Tensor).data[0] as number;\n    return probability;\n  }\n\n  /** Start VAD processing on a MediaStream (typically from getUserMedia). */\n  start(stream: MediaStream): void {\n    if (this._isDestroyed) {\n      throw new Error(`${LOG_PREFIX} Cannot start after destroy`);\n    }\n    if (!this._isReady) {\n      throw new Error(`${LOG_PREFIX} Model not loaded. Call init() first.`);\n    }\n    if (this._isStarted) {\n      this._log('Already started — stopping previous session first');\n      this.stop();\n    }\n\n    this._log('Starting VAD on MediaStream');\n    this._stream = stream;\n    this._isStarted = true;\n\n    // Reset speech tracking state.\n    this._isSpeaking = false;\n    this._consecutiveSpeechFrames = 0;\n    this._consecutiveSilenceFrames = 0;\n    this._frameBuffer = new Float32Array(FRAME_SIZE);\n    this._frameBufferOffset = 0;\n\n    // Reset LSTM states for a fresh stream.\n    this._resetStates();\n\n    // Begin noise floor calibration.\n    this._isCalibrating = true;\n    this._calibrationSamples = [];\n    this._calibrationFramesNeeded = Math.max(1, Math.floor(\n      (CALIBRATION_DURATION_MS / 1000) * this._sampleRate / FRAME_SIZE,\n    ));\n    this._log('Calibrating noise floor for', this._calibrationFramesNeeded, 'frames');\n\n    // Build the audio processing pipeline.\n    this._setupAudioPipeline(stream);\n  }\n\n  /** Stop VAD processing and release audio resources (but keep the model). */\n  stop(): void {\n    if (!this._isStarted) return;\n\n    this._log('Stopping VAD');\n\n    // Tear down audio nodes.\n    this._teardownAudioPipeline();\n\n    // If we were speaking, emit speech-end.\n    if (this._isSpeaking) {\n      this._isSpeaking = false;\n      this._emit({\n        type: 'speech-end',\n        timestamp: Date.now(),\n        probability: 0,\n      });\n    }\n\n    // Reset state.\n    this._isStarted = false;\n    this._isSpeaking = false;\n    this._consecutiveSpeechFrames = 0;\n    this._consecutiveSilenceFrames = 0;\n    this._frameBufferOffset = 0;\n    this._isCalibrating = false;\n    this._calibrationSamples = [];\n    this._processingPromise = Promise.resolve();\n\n    this._log('VAD stopped');\n  }\n\n  /** Register a callback for speech-start events. Returns an unsubscribe function. */\n  onSpeechStart(callback: VADCallback): () => void {\n    return this._on('speech-start', callback);\n  }\n\n  /** Register a callback for speech-end events. Returns an unsubscribe function. */\n  onSpeechEnd(callback: VADCallback): () => void {\n    return this._on('speech-end', callback);\n  }\n\n  /** Register a callback for vad-ready events. Returns an unsubscribe function. */\n  onReady(callback: VADCallback): () => void {\n    return this._on('vad-ready', callback);\n  }\n\n  /** Whether the ONNX model is loaded and ready. */\n  get isReady(): boolean {\n    return this._isReady;\n  }\n\n  /** Whether speech is currently detected. */\n  get isSpeaking(): boolean {\n    return this._isSpeaking;\n  }\n\n  /** The MediaStream currently being processed, or null. */\n  get stream(): MediaStream | null {\n    return this._stream;\n  }\n\n  /** Release ONNX model session and all audio resources. */\n  async destroy(): Promise<void> {\n    if (this._isDestroyed) return;\n    this._log('Destroying...');\n\n    this.stop();\n\n    if (this._session) {\n      await this._session.release();\n      this._session = null;\n    }\n\n    // Dispose tensors.\n    this._h?.dispose();\n    this._c?.dispose();\n    this._h = null;\n    this._c = null;\n\n    this._isReady = false;\n    this._isDestroyed = true;\n    this._listeners.clear();\n\n    this._log('Destroyed');\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: Event system\n  // -------------------------------------------------------------------------\n\n  private _on(type: VADEventType, callback: VADCallback): () => void {\n    let set = this._listeners.get(type);\n    if (!set) {\n      set = new Set();\n      this._listeners.set(type, set);\n    }\n    set.add(callback);\n    return () => {\n      set!.delete(callback);\n    };\n  }\n\n  private _emit(event: VADEvent): void {\n    const set = this._listeners.get(event.type);\n    if (!set) return;\n    for (const cb of set) {\n      try {\n        cb(event);\n      } catch (err) {\n        // eslint-disable-next-line no-console\n        console.error(`${LOG_PREFIX} Error in ${event.type} callback:`, err);\n      }\n    }\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: Audio pipeline\n  // -------------------------------------------------------------------------\n\n  private _setupAudioPipeline(stream: MediaStream): void {\n    // Determine the incoming sample rate.\n    const tracks = stream.getAudioTracks();\n    const trackSettings = tracks[0]?.getSettings();\n    const inputSampleRate = trackSettings?.sampleRate ?? 48000;\n\n    this._log('Input sample rate:', inputSampleRate);\n\n    // Create AudioContext at the input sample rate so we don't double-resample.\n    // SSR guard: AudioContext may not exist.\n    if (typeof AudioContext === 'undefined' && typeof webkitAudioContext === 'undefined') {\n      throw new Error(`${LOG_PREFIX} AudioContext is not available in this environment`);\n    }\n\n    const AudioContextClass =\n      typeof AudioContext !== 'undefined'\n        ? AudioContext\n        : // eslint-disable-next-line @typescript-eslint/no-explicit-any\n          (globalThis as any).webkitAudioContext as typeof AudioContext;\n\n    this._audioContext = new AudioContextClass({ sampleRate: inputSampleRate });\n    this._ownsAudioContext = true;\n\n    this._sourceNode = this._audioContext.createMediaStreamSource(stream);\n\n    // Try AudioWorklet first, fall back to ScriptProcessorNode.\n    this._setupScriptProcessor(inputSampleRate);\n  }\n\n  /**\n   * ScriptProcessorNode fallback (works everywhere, including Safari).\n   * We use a buffer size of 4096 which gives ~85 ms of audio at 48 kHz.\n   */\n  private _setupScriptProcessor(inputSampleRate: number): void {\n    if (!this._audioContext || !this._sourceNode) return;\n\n    // Buffer size must be a power of 2: 256, 512, 1024, 2048, 4096, 8192, 16384.\n    const bufferSize = 4096;\n    const processor = this._audioContext.createScriptProcessor(bufferSize, 1, 1);\n\n    processor.onaudioprocess = (event: AudioProcessingEvent) => {\n      if (!this._isStarted) return;\n\n      const inputData = event.inputBuffer.getChannelData(0);\n\n      // Resample to target rate if needed.\n      const resampled =\n        inputSampleRate !== this._sampleRate\n          ? resample(inputData, inputSampleRate, this._sampleRate)\n          : new Float32Array(inputData);\n\n      // Feed resampled audio into frame-sized chunks.\n      this._feedAudio(resampled);\n    };\n\n    this._sourceNode.connect(processor);\n    processor.connect(this._audioContext.destination);\n    this._workletNode = processor;\n\n    this._log('Audio pipeline set up (ScriptProcessorNode)');\n  }\n\n  /**\n   * Accumulate resampled audio into FRAME_SIZE chunks and process each full frame.\n   */\n  private _feedAudio(samples: Float32Array): void {\n    let offset = 0;\n\n    while (offset < samples.length) {\n      const remaining = FRAME_SIZE - this._frameBufferOffset;\n      const available = samples.length - offset;\n      const toCopy = Math.min(remaining, available);\n\n      this._frameBuffer.set(\n        samples.subarray(offset, offset + toCopy),\n        this._frameBufferOffset,\n      );\n      this._frameBufferOffset += toCopy;\n      offset += toCopy;\n\n      if (this._frameBufferOffset === FRAME_SIZE) {\n        const frame = new Float32Array(this._frameBuffer);\n        this._frameBufferOffset = 0;\n\n        // Serialise inference calls to avoid overlapping ONNX sessions.\n        this._processingPromise = this._processingPromise.then(() =>\n          this._handleFrame(frame),\n        );\n      }\n    }\n  }\n\n  /**\n   * Process a single FRAME_SIZE frame: run inference and update speech state.\n   */\n  private async _handleFrame(frame: Float32Array): Promise<void> {\n    if (!this._isStarted || !this._session) return;\n\n    let probability: number;\n    try {\n      probability = await this.processFrame(frame);\n    } catch (err) {\n      if (this._debug) {\n        // eslint-disable-next-line no-console\n        console.error(`${LOG_PREFIX} Inference error:`, err);\n      }\n      return;\n    }\n\n    // Noise floor calibration phase.\n    if (this._isCalibrating) {\n      this._calibrationSamples.push(probability);\n\n      if (this._calibrationSamples.length >= this._calibrationFramesNeeded) {\n        this._finishCalibration();\n      }\n      return;\n    }\n\n    // Speech state machine.\n    const isSpeechFrame = probability >= this._calibratedThreshold;\n\n    if (isSpeechFrame) {\n      this._consecutiveSpeechFrames++;\n      this._consecutiveSilenceFrames = 0;\n\n      if (!this._isSpeaking && this._consecutiveSpeechFrames >= this._minSpeechFrames) {\n        this._isSpeaking = true;\n        this._log('Speech started, probability:', probability.toFixed(3));\n        this._emit({\n          type: 'speech-start',\n          timestamp: Date.now(),\n          probability,\n        });\n      }\n    } else {\n      this._consecutiveSilenceFrames++;\n      // Do NOT reset _consecutiveSpeechFrames here — only reset when speech-end fires.\n\n      if (this._isSpeaking && this._consecutiveSilenceFrames >= this._silenceFrames) {\n        this._isSpeaking = false;\n        this._consecutiveSpeechFrames = 0;\n        this._log('Speech ended, probability:', probability.toFixed(3));\n        this._emit({\n          type: 'speech-end',\n          timestamp: Date.now(),\n          probability,\n        });\n      }\n    }\n  }\n\n  private _finishCalibration(): void {\n    if (this._calibrationSamples.length === 0) {\n      this._isCalibrating = false;\n      return;\n    }\n\n    // Compute average noise floor probability.\n    const sum = this._calibrationSamples.reduce((a, b) => a + b, 0);\n    const avgNoise = sum / this._calibrationSamples.length;\n\n    // If the ambient noise floor is high, nudge the threshold above it.\n    // We add a margin so we don't constantly trigger on background noise.\n    const NOISE_MARGIN = 0.15;\n    if (avgNoise + NOISE_MARGIN > this._threshold) {\n      this._calibratedThreshold = Math.min(avgNoise + NOISE_MARGIN, 0.95);\n      this._log(\n        'Noise floor is high. Adjusted threshold from',\n        this._threshold.toFixed(3),\n        'to',\n        this._calibratedThreshold.toFixed(3),\n        '(avg noise:',\n        avgNoise.toFixed(3) + ')',\n      );\n    } else {\n      this._calibratedThreshold = this._threshold;\n      this._log('Noise floor OK, avg:', avgNoise.toFixed(3), '— keeping threshold at', this._threshold.toFixed(3));\n    }\n\n    // Recompute frame counters in case threshold changed min speech behaviour.\n    this._minSpeechFrames = Math.ceil(this._minSpeechDurationMs / this._frameDurationMs);\n    this._silenceFrames = Math.ceil(this._silenceDurationMs / this._frameDurationMs);\n\n    this._isCalibrating = false;\n    this._calibrationSamples = [];\n  }\n\n  private _teardownAudioPipeline(): void {\n    if (this._workletNode) {\n      try {\n        this._workletNode.disconnect();\n      } catch {\n        // Ignore disconnect errors.\n      }\n      if ('onaudioprocess' in this._workletNode) {\n        (this._workletNode as ScriptProcessorNode).onaudioprocess = null;\n      }\n      this._workletNode = null;\n    }\n\n    if (this._sourceNode) {\n      try {\n        this._sourceNode.disconnect();\n      } catch {\n        // Ignore.\n      }\n      this._sourceNode = null;\n    }\n\n    if (this._audioContext && this._ownsAudioContext) {\n      try {\n        void this._audioContext.close();\n      } catch {\n        // Ignore.\n      }\n      this._audioContext = null;\n      this._ownsAudioContext = false;\n    }\n\n    this._stream = null;\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: ONNX state helpers\n  // -------------------------------------------------------------------------\n\n  /** Reset the LSTM hidden and cell states to zeros. */\n  private _resetStates(): void {\n    // Dispose any existing tensors to free memory.\n    this._h?.dispose();\n    this._c?.dispose();\n\n    const zeros = new Float32Array(2 * STATE_SIZE).fill(0);\n    this._h = new ort.Tensor('float32', zeros.slice(0, STATE_SIZE), [2, 1, 64]);\n    this._c = new ort.Tensor('float32', zeros.slice(STATE_SIZE), [2, 1, 64]);\n  }\n\n  // -------------------------------------------------------------------------\n  // Private: Logging\n  // -------------------------------------------------------------------------\n\n  private _log(...args: unknown[]): void {\n    if (!this._debug) return;\n    // eslint-disable-next-line no-console\n    console.log(LOG_PREFIX, ...args);\n  }\n}\n\n// ---------------------------------------------------------------------------\n// Convenience factory\n// ---------------------------------------------------------------------------\n\n/**\n * Create and initialise a SileroVAD instance in one call.\n *\n * ```ts\n * const vad = await createVAD({ debug: true });\n * vad.onSpeechStart(() => console.log('speaking'));\n * vad.start(stream);\n * ```\n */\nexport async function createVAD(options?: VADOptions): Promise<SileroVAD> {\n  const vad = new SileroVAD(options);\n  await vad.init();\n  return vad;\n}\n\n// Re-export the frame size constant so consumers can align their buffers.\nexport { FRAME_SIZE, TARGET_SAMPLE_RATE };\n\n// Type-only declaration for environments that provide webkitAudioContext.\ndeclare global {\n  // eslint-disable-next-line no-var\n  var webkitAudioContext: typeof AudioContext | undefined;\n}\n"],"mappings":";AACA,YAAY,SAAS;AAEd,IAAM,cAAc;AAM3B,IAAM,aAAa;AAGnB,IAAM,oBACJ;AAGF,IAAM,aAAa,iBAAiB,WAAW;AAC/C,IAAM,kBAAkB;AAGxB,IAAM,aAAa;AAGnB,IAAM,qBAAqB;AAG3B,IAAM,0BAA0B;AAGhC,IAAM,aAAa;AAmCnB,eAAe,qBAAkD;AAC/D,MAAI,OAAO,WAAW,YAAa,QAAO;AAC1C,MAAI;AACF,UAAM,QAAQ,MAAM,OAAO,KAAK,UAAU;AAC1C,UAAM,WAAW,MAAM,MAAM,MAAM,eAAe;AAClD,WAAO,WAAW,SAAS,YAAY,IAAI;AAAA,EAC7C,QAAQ;AAEN,WAAO;AAAA,EACT;AACF;AAEA,eAAe,iBAAiB,MAAkC;AAChE,MAAI,OAAO,WAAW,YAAa;AACnC,MAAI;AACF,UAAM,QAAQ,MAAM,OAAO,KAAK,UAAU;AAC1C,UAAM,MAAM,IAAI,iBAAiB,IAAI,SAAS,IAAI,CAAC;AAAA,EACrD,QAAQ;AAAA,EAER;AACF;AAUA,SAAS,SACP,OACA,WACA,YACc;AACd,MAAI,cAAc,WAAY,QAAO;AACrC,QAAM,QAAQ,YAAY;AAC1B,QAAM,eAAe,KAAK,MAAM,MAAM,SAAS,KAAK;AACpD,QAAM,SAAS,IAAI,aAAa,YAAY;AAC5C,WAAS,IAAI,GAAG,IAAI,cAAc,KAAK;AACrC,UAAM,WAAW,IAAI;AACrB,UAAM,WAAW,KAAK,MAAM,QAAQ;AACpC,UAAM,UAAU,KAAK,IAAI,WAAW,GAAG,MAAM,SAAS,CAAC;AACvD,UAAM,OAAO,WAAW;AACxB,WAAO,CAAC,IAAK,MAAM,QAAQ,KAAgB,IAAI,QAAS,MAAM,OAAO,IAAe;AAAA,EACtF;AACA,SAAO;AACT;AAMO,IAAM,YAAN,MAAgB;AAAA;AAAA,EAEJ;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA;AAAA,EAGT,WAAwC;AAAA,EACxC,KAAwB;AAAA,EACxB,KAAwB;AAAA;AAAA,EAGxB,gBAAqC;AAAA,EACrC,oBAAoB;AAAA,EACpB,cAAiD;AAAA,EACjD,eAA8D;AAAA,EAC9D,UAA8B;AAAA;AAAA,EAG9B,eAA6B,IAAI,aAAa,CAAC;AAAA,EAC/C,qBAAqB;AAAA;AAAA,EAGrB,WAAW;AAAA,EACX,cAAc;AAAA,EACd,aAAa;AAAA,EACb,eAAe;AAAA;AAAA,EAGf,2BAA2B;AAAA,EAC3B,4BAA4B;AAAA,EAC5B;AAAA,EACA;AAAA,EACA;AAAA;AAAA,EAGA,iBAAiB;AAAA,EACjB,sBAAgC,CAAC;AAAA,EACjC,2BAA2B;AAAA,EAC3B;AAAA;AAAA,EAGA,aAAkD,oBAAI,IAAI;AAAA;AAAA,EAG1D,qBAAoC,QAAQ,QAAQ;AAAA,EAE5D,YAAY,SAAsB;AAChC,SAAK,aAAa,SAAS,aAAa;AACxC,SAAK,uBAAuB,SAAS,uBAAuB;AAC5D,SAAK,qBAAqB,SAAS,qBAAqB;AACxD,SAAK,cAAc,SAAS,cAAc;AAC1C,SAAK,SAAS,SAAS,SAAS;AAChC,SAAK,YAAY,SAAS,YAAY;AACtC,SAAK,uBAAuB,KAAK;AAGjC,SAAK,mBAAoB,aAAa,KAAK,cAAe;AAC1D,SAAK,mBAAmB,KAAK,KAAK,KAAK,uBAAuB,KAAK,gBAAgB;AACnF,SAAK,iBAAiB,KAAK,KAAK,KAAK,qBAAqB,KAAK,gBAAgB;AAE/E,SAAK,KAAK,wBAAwB;AAAA,MAChC,WAAW,KAAK;AAAA,MAChB,qBAAqB,KAAK;AAAA,MAC1B,mBAAmB,KAAK;AAAA,MACxB,YAAY,KAAK;AAAA,MACjB,UAAU,KAAK;AAAA,IACjB,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,OAAsB;AAC1B,QAAI,KAAK,cAAc;AACrB,YAAM,IAAI,MAAM,GAAG,UAAU,4BAA4B;AAAA,IAC3D;AACA,QAAI,KAAK,UAAU;AACjB,WAAK,KAAK,qCAAgC;AAC1C;AAAA,IACF;AAEA,SAAK,KAAK,iBAAiB;AAG3B,QAAI,cAAc,MAAM,mBAAmB;AAC3C,QAAI,aAAa;AACf,WAAK,KAAK,6BAA6B;AAAA,IACzC,OAAO;AACL,WAAK,KAAK,uBAAuB,KAAK,SAAS;AAC/C,YAAM,WAAW,MAAM,MAAM,KAAK,SAAS;AAC3C,UAAI,CAAC,SAAS,IAAI;AAChB,cAAM,IAAI;AAAA,UACR,GAAG,UAAU,2BAA2B,SAAS,MAAM,IAAI,SAAS,UAAU;AAAA,QAChF;AAAA,MACF;AACA,oBAAc,MAAM,SAAS,YAAY;AACzC,WAAK,KAAK,wBAAwB,YAAY,YAAY,OAAO;AAGjE,YAAM,iBAAiB,WAAW;AAClC,WAAK,KAAK,0BAA0B;AAAA,IACtC;AAGA,SAAK,WAAW,MAAU,qBAAiB,OAAO,aAAa;AAAA,MAC7D,oBAAoB,CAAC,MAAM;AAAA,MAC3B,wBAAwB;AAAA,IAC1B,CAAC;AAGD,SAAK,aAAa;AAElB,SAAK,WAAW;AAChB,SAAK,KAAK,wBAAwB;AAElC,SAAK,MAAM;AAAA,MACT,MAAM;AAAA,MACN,WAAW,KAAK,IAAI;AAAA,IACtB,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,aAAa,WAA0C;AAC3D,QAAI,CAAC,KAAK,UAAU;AAClB,YAAM,IAAI,MAAM,GAAG,UAAU,uCAAuC;AAAA,IACtE;AACA,QAAI,UAAU,WAAW,YAAY;AACnC,YAAM,IAAI;AAAA,QACR,GAAG,UAAU,aAAa,UAAU,iBAAiB,UAAU,MAAM;AAAA,MACvE;AAAA,IACF;AAEA,UAAM,cAAc,IAAQ,WAAO,WAAW,WAAW,CAAC,GAAG,UAAU,CAAC;AACxE,UAAM,WAAW,IAAQ,WAAO,SAAS,cAAc,KAAK,CAAC,OAAO,KAAK,WAAW,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;AAE5F,UAAM,QAAoC;AAAA,MACxC,OAAO;AAAA,MACP,IAAI;AAAA,MACJ,GAAG,KAAK;AAAA,MACR,GAAG,KAAK;AAAA,IACV;AAEA,UAAM,UAAU,MAAM,KAAK,SAAS,IAAI,KAAK;AAG7C,SAAK,KAAK,QAAQ,IAAI;AACtB,SAAK,KAAK,QAAQ,IAAI;AAEtB,UAAM,cAAe,QAAQ,QAAQ,EAAiB,KAAK,CAAC;AAC5D,WAAO;AAAA,EACT;AAAA;AAAA,EAGA,MAAM,QAA2B;AAC/B,QAAI,KAAK,cAAc;AACrB,YAAM,IAAI,MAAM,GAAG,UAAU,6BAA6B;AAAA,IAC5D;AACA,QAAI,CAAC,KAAK,UAAU;AAClB,YAAM,IAAI,MAAM,GAAG,UAAU,uCAAuC;AAAA,IACtE;AACA,QAAI,KAAK,YAAY;AACnB,WAAK,KAAK,wDAAmD;AAC7D,WAAK,KAAK;AAAA,IACZ;AAEA,SAAK,KAAK,6BAA6B;AACvC,SAAK,UAAU;AACf,SAAK,aAAa;AAGlB,SAAK,cAAc;AACnB,SAAK,2BAA2B;AAChC,SAAK,4BAA4B;AACjC,SAAK,eAAe,IAAI,aAAa,UAAU;AAC/C,SAAK,qBAAqB;AAG1B,SAAK,aAAa;AAGlB,SAAK,iBAAiB;AACtB,SAAK,sBAAsB,CAAC;AAC5B,SAAK,2BAA2B,KAAK,IAAI,GAAG,KAAK;AAAA,MAC9C,0BAA0B,MAAQ,KAAK,cAAc;AAAA,IACxD,CAAC;AACD,SAAK,KAAK,+BAA+B,KAAK,0BAA0B,QAAQ;AAGhF,SAAK,oBAAoB,MAAM;AAAA,EACjC;AAAA;AAAA,EAGA,OAAa;AACX,QAAI,CAAC,KAAK,WAAY;AAEtB,SAAK,KAAK,cAAc;AAGxB,SAAK,uBAAuB;AAG5B,QAAI,KAAK,aAAa;AACpB,WAAK,cAAc;AACnB,WAAK,MAAM;AAAA,QACT,MAAM;AAAA,QACN,WAAW,KAAK,IAAI;AAAA,QACpB,aAAa;AAAA,MACf,CAAC;AAAA,IACH;AAGA,SAAK,aAAa;AAClB,SAAK,cAAc;AACnB,SAAK,2BAA2B;AAChC,SAAK,4BAA4B;AACjC,SAAK,qBAAqB;AAC1B,SAAK,iBAAiB;AACtB,SAAK,sBAAsB,CAAC;AAC5B,SAAK,qBAAqB,QAAQ,QAAQ;AAE1C,SAAK,KAAK,aAAa;AAAA,EACzB;AAAA;AAAA,EAGA,cAAc,UAAmC;AAC/C,WAAO,KAAK,IAAI,gBAAgB,QAAQ;AAAA,EAC1C;AAAA;AAAA,EAGA,YAAY,UAAmC;AAC7C,WAAO,KAAK,IAAI,cAAc,QAAQ;AAAA,EACxC;AAAA;AAAA,EAGA,QAAQ,UAAmC;AACzC,WAAO,KAAK,IAAI,aAAa,QAAQ;AAAA,EACvC;AAAA;AAAA,EAGA,IAAI,UAAmB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,aAAsB;AACxB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,SAA6B;AAC/B,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,MAAM,UAAyB;AAC7B,QAAI,KAAK,aAAc;AACvB,SAAK,KAAK,eAAe;AAEzB,SAAK,KAAK;AAEV,QAAI,KAAK,UAAU;AACjB,YAAM,KAAK,SAAS,QAAQ;AAC5B,WAAK,WAAW;AAAA,IAClB;AAGA,SAAK,IAAI,QAAQ;AACjB,SAAK,IAAI,QAAQ;AACjB,SAAK,KAAK;AACV,SAAK,KAAK;AAEV,SAAK,WAAW;AAChB,SAAK,eAAe;AACpB,SAAK,WAAW,MAAM;AAEtB,SAAK,KAAK,WAAW;AAAA,EACvB;AAAA;AAAA;AAAA;AAAA,EAMQ,IAAI,MAAoB,UAAmC;AACjE,QAAI,MAAM,KAAK,WAAW,IAAI,IAAI;AAClC,QAAI,CAAC,KAAK;AACR,YAAM,oBAAI,IAAI;AACd,WAAK,WAAW,IAAI,MAAM,GAAG;AAAA,IAC/B;AACA,QAAI,IAAI,QAAQ;AAChB,WAAO,MAAM;AACX,UAAK,OAAO,QAAQ;AAAA,IACtB;AAAA,EACF;AAAA,EAEQ,MAAM,OAAuB;AACnC,UAAM,MAAM,KAAK,WAAW,IAAI,MAAM,IAAI;AAC1C,QAAI,CAAC,IAAK;AACV,eAAW,MAAM,KAAK;AACpB,UAAI;AACF,WAAG,KAAK;AAAA,MACV,SAAS,KAAK;AAEZ,gBAAQ,MAAM,GAAG,UAAU,aAAa,MAAM,IAAI,cAAc,GAAG;AAAA,MACrE;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAMQ,oBAAoB,QAA2B;AAErD,UAAM,SAAS,OAAO,eAAe;AACrC,UAAM,gBAAgB,OAAO,CAAC,GAAG,YAAY;AAC7C,UAAM,kBAAkB,eAAe,cAAc;AAErD,SAAK,KAAK,sBAAsB,eAAe;AAI/C,QAAI,OAAO,iBAAiB,eAAe,OAAO,uBAAuB,aAAa;AACpF,YAAM,IAAI,MAAM,GAAG,UAAU,oDAAoD;AAAA,IACnF;AAEA,UAAM,oBACJ,OAAO,iBAAiB,cACpB;AAAA;AAAA,MAEC,WAAmB;AAAA;AAE1B,SAAK,gBAAgB,IAAI,kBAAkB,EAAE,YAAY,gBAAgB,CAAC;AAC1E,SAAK,oBAAoB;AAEzB,SAAK,cAAc,KAAK,cAAc,wBAAwB,MAAM;AAGpE,SAAK,sBAAsB,eAAe;AAAA,EAC5C;AAAA;AAAA;AAAA;AAAA;AAAA,EAMQ,sBAAsB,iBAA+B;AAC3D,QAAI,CAAC,KAAK,iBAAiB,CAAC,KAAK,YAAa;AAG9C,UAAM,aAAa;AACnB,UAAM,YAAY,KAAK,cAAc,sBAAsB,YAAY,GAAG,CAAC;AAE3E,cAAU,iBAAiB,CAAC,UAAgC;AAC1D,UAAI,CAAC,KAAK,WAAY;AAEtB,YAAM,YAAY,MAAM,YAAY,eAAe,CAAC;AAGpD,YAAM,YACJ,oBAAoB,KAAK,cACrB,SAAS,WAAW,iBAAiB,KAAK,WAAW,IACrD,IAAI,aAAa,SAAS;AAGhC,WAAK,WAAW,SAAS;AAAA,IAC3B;AAEA,SAAK,YAAY,QAAQ,SAAS;AAClC,cAAU,QAAQ,KAAK,cAAc,WAAW;AAChD,SAAK,eAAe;AAEpB,SAAK,KAAK,6CAA6C;AAAA,EACzD;AAAA;AAAA;AAAA;AAAA,EAKQ,WAAW,SAA6B;AAC9C,QAAI,SAAS;AAEb,WAAO,SAAS,QAAQ,QAAQ;AAC9B,YAAM,YAAY,aAAa,KAAK;AACpC,YAAM,YAAY,QAAQ,SAAS;AACnC,YAAM,SAAS,KAAK,IAAI,WAAW,SAAS;AAE5C,WAAK,aAAa;AAAA,QAChB,QAAQ,SAAS,QAAQ,SAAS,MAAM;AAAA,QACxC,KAAK;AAAA,MACP;AACA,WAAK,sBAAsB;AAC3B,gBAAU;AAEV,UAAI,KAAK,uBAAuB,YAAY;AAC1C,cAAM,QAAQ,IAAI,aAAa,KAAK,YAAY;AAChD,aAAK,qBAAqB;AAG1B,aAAK,qBAAqB,KAAK,mBAAmB;AAAA,UAAK,MACrD,KAAK,aAAa,KAAK;AAAA,QACzB;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,MAAc,aAAa,OAAoC;AAC7D,QAAI,CAAC,KAAK,cAAc,CAAC,KAAK,SAAU;AAExC,QAAI;AACJ,QAAI;AACF,oBAAc,MAAM,KAAK,aAAa,KAAK;AAAA,IAC7C,SAAS,KAAK;AACZ,UAAI,KAAK,QAAQ;AAEf,gBAAQ,MAAM,GAAG,UAAU,qBAAqB,GAAG;AAAA,MACrD;AACA;AAAA,IACF;AAGA,QAAI,KAAK,gBAAgB;AACvB,WAAK,oBAAoB,KAAK,WAAW;AAEzC,UAAI,KAAK,oBAAoB,UAAU,KAAK,0BAA0B;AACpE,aAAK,mBAAmB;AAAA,MAC1B;AACA;AAAA,IACF;AAGA,UAAM,gBAAgB,eAAe,KAAK;AAE1C,QAAI,eAAe;AACjB,WAAK;AACL,WAAK,4BAA4B;AAEjC,UAAI,CAAC,KAAK,eAAe,KAAK,4BAA4B,KAAK,kBAAkB;AAC/E,aAAK,cAAc;AACnB,aAAK,KAAK,gCAAgC,YAAY,QAAQ,CAAC,CAAC;AAChE,aAAK,MAAM;AAAA,UACT,MAAM;AAAA,UACN,WAAW,KAAK,IAAI;AAAA,UACpB;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF,OAAO;AACL,WAAK;AAGL,UAAI,KAAK,eAAe,KAAK,6BAA6B,KAAK,gBAAgB;AAC7E,aAAK,cAAc;AACnB,aAAK,2BAA2B;AAChC,aAAK,KAAK,8BAA8B,YAAY,QAAQ,CAAC,CAAC;AAC9D,aAAK,MAAM;AAAA,UACT,MAAM;AAAA,UACN,WAAW,KAAK,IAAI;AAAA,UACpB;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,qBAA2B;AACjC,QAAI,KAAK,oBAAoB,WAAW,GAAG;AACzC,WAAK,iBAAiB;AACtB;AAAA,IACF;AAGA,UAAM,MAAM,KAAK,oBAAoB,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC9D,UAAM,WAAW,MAAM,KAAK,oBAAoB;AAIhD,UAAM,eAAe;AACrB,QAAI,WAAW,eAAe,KAAK,YAAY;AAC7C,WAAK,uBAAuB,KAAK,IAAI,WAAW,cAAc,IAAI;AAClE,WAAK;AAAA,QACH;AAAA,QACA,KAAK,WAAW,QAAQ,CAAC;AAAA,QACzB;AAAA,QACA,KAAK,qBAAqB,QAAQ,CAAC;AAAA,QACnC;AAAA,QACA,SAAS,QAAQ,CAAC,IAAI;AAAA,MACxB;AAAA,IACF,OAAO;AACL,WAAK,uBAAuB,KAAK;AACjC,WAAK,KAAK,wBAAwB,SAAS,QAAQ,CAAC,GAAG,+BAA0B,KAAK,WAAW,QAAQ,CAAC,CAAC;AAAA,IAC7G;AAGA,SAAK,mBAAmB,KAAK,KAAK,KAAK,uBAAuB,KAAK,gBAAgB;AACnF,SAAK,iBAAiB,KAAK,KAAK,KAAK,qBAAqB,KAAK,gBAAgB;AAE/E,SAAK,iBAAiB;AACtB,SAAK,sBAAsB,CAAC;AAAA,EAC9B;AAAA,EAEQ,yBAA+B;AACrC,QAAI,KAAK,cAAc;AACrB,UAAI;AACF,aAAK,aAAa,WAAW;AAAA,MAC/B,QAAQ;AAAA,MAER;AACA,UAAI,oBAAoB,KAAK,cAAc;AACzC,QAAC,KAAK,aAAqC,iBAAiB;AAAA,MAC9D;AACA,WAAK,eAAe;AAAA,IACtB;AAEA,QAAI,KAAK,aAAa;AACpB,UAAI;AACF,aAAK,YAAY,WAAW;AAAA,MAC9B,QAAQ;AAAA,MAER;AACA,WAAK,cAAc;AAAA,IACrB;AAEA,QAAI,KAAK,iBAAiB,KAAK,mBAAmB;AAChD,UAAI;AACF,aAAK,KAAK,cAAc,MAAM;AAAA,MAChC,QAAQ;AAAA,MAER;AACA,WAAK,gBAAgB;AACrB,WAAK,oBAAoB;AAAA,IAC3B;AAEA,SAAK,UAAU;AAAA,EACjB;AAAA;AAAA;AAAA;AAAA;AAAA,EAOQ,eAAqB;AAE3B,SAAK,IAAI,QAAQ;AACjB,SAAK,IAAI,QAAQ;AAEjB,UAAM,QAAQ,IAAI,aAAa,IAAI,UAAU,EAAE,KAAK,CAAC;AACrD,SAAK,KAAK,IAAQ,WAAO,WAAW,MAAM,MAAM,GAAG,UAAU,GAAG,CAAC,GAAG,GAAG,EAAE,CAAC;AAC1E,SAAK,KAAK,IAAQ,WAAO,WAAW,MAAM,MAAM,UAAU,GAAG,CAAC,GAAG,GAAG,EAAE,CAAC;AAAA,EACzE;AAAA;AAAA;AAAA;AAAA,EAMQ,QAAQ,MAAuB;AACrC,QAAI,CAAC,KAAK,OAAQ;AAElB,YAAQ,IAAI,YAAY,GAAG,IAAI;AAAA,EACjC;AACF;AAeA,eAAsB,UAAU,SAA0C;AACxE,QAAM,MAAM,IAAI,UAAU,OAAO;AACjC,QAAM,IAAI,KAAK;AACf,SAAO;AACT;","names":[]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@guidekit/vad",
-  "version": "0.1.0-beta.1",
+  "version": "0.1.0-beta.2",
   "description": "Silero VAD model for GuideKit SDK — voice activity detection via ONNX Runtime",
   "type": "module",
   "main": "./dist/index.cjs",
@@ -16,17 +16,9 @@
   "files": [
     "dist",
     "models",
-    "../../LICENSE",
+    "LICENSE",
     "README.md"
   ],
-  "scripts": {
-    "build": "tsup",
-    "dev": "tsup --watch",
-    "test": "echo 'No tests yet'",
-    "test:unit": "echo 'No tests yet'",
-    "typecheck": "tsc --noEmit",
-    "clean": "rm -rf dist"
-  },
   "dependencies": {
     "onnxruntime-web": "^1.21.0"
   },
@@ -36,5 +28,34 @@
     "vitest": "^3.0.0"
   },
   "sideEffects": false,
-  "license": "MIT"
-}
+  "license": "MIT",
+  "author": "Riazul Islam",
+  "keywords": [
+    "guidekit",
+    "vad",
+    "voice-activity-detection",
+    "silero",
+    "onnx",
+    "audio"
+  ],
+  "engines": {
+    "node": ">=18"
+  },
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/riaz37/guidekit.git",
+    "directory": "packages/vad"
+  },
+  "homepage": "https://github.com/riaz37/guidekit#readme",
+  "bugs": {
+    "url": "https://github.com/riaz37/guidekit/issues"
+  },
+  "scripts": {
+    "build": "tsup",
+    "dev": "tsup --watch",
+    "test": "echo 'No tests yet'",
+    "test:unit": "echo 'No tests yet'",
+    "typecheck": "tsc --noEmit",
+    "clean": "rm -rf dist"
+  }
+}