@tryhamster/gerbil 1.0.0-rc.6 → 1.0.0-rc.7

package/dist/{stt-Te8Qz-Ay.js.map → stt-Bu-E23Sc.js.map}

@@ -1 +1 @@
-{"version":3,"file":"stt-Te8Qz-Ay.js","names":["WHISPER_MODELS: STTModelConfig[]","tfDevice: \"webgpu\" | \"cpu\" | \"wasm\"","audioData: Float32Array","pipelineOptions: any","transcribeResult: TranscribeResult","audioBuffer: Float32Array[]","intervalId: ReturnType<typeof setInterval> | null","e: any"],"sources":["../src/core/stt.ts"],"sourcesContent":["/**\n * Speech-to-Text with Whisper\n *\n * Provides local speech recognition using Whisper ONNX models via transformers.js.\n * Supports multiple model sizes and languages.\n *\n * @example\n * ```ts\n * const stt = new WhisperSTT();\n * await stt.load({ onProgress: (p) => console.log(p.status) });\n *\n * // Transcribe audio (Float32Array at 16kHz)\n * const result = await stt.transcribe(audioData);\n * console.log(result.text);\n *\n * // With timestamps\n * const result = await stt.transcribe(audioData, { timestamps: true });\n * for (const seg of result.segments) {\n *   console.log(`[${seg.start.toFixed(1)}s] ${seg.text}`);\n * }\n * ```\n */\n\nimport type {\n  LoadSTTOptions,\n  ProgressInfo,\n  STTModelConfig,\n  StreamingTranscriptionOptions,\n  StreamingTranscriptionSession,\n  TranscribeOptions,\n  TranscribeResult,\n  TranscribeSegment,\n} from \"./types.js\";\n\n// ============================================\n// Model Registry\n// ============================================\n\n/**\n * Available Whisper models\n * Ordered by size (smallest first)\n */\nexport const WHISPER_MODELS: STTModelConfig[] = [\n  {\n    id: \"whisper-tiny.en\",\n    repo: \"onnx-community/whisper-tiny.en\",\n    description: \"Tiny English-only model, fastest\",\n    size: \"39M\",\n    multilingual: false,\n    languages: [\"en\"],\n    sampleRate: 16000,\n  },\n  {\n    id: \"whisper-tiny\",\n    repo: \"onnx-community/whisper-tiny\",\n    description: \"Tiny multilingual model\",\n    size: \"39M\",\n    multilingual: true,\n    languages: [\"en\", \"es\", \"fr\", \"de\", \"it\", \"pt\", \"nl\", \"ru\", \"zh\", \"ja\", \"ko\"],\n    sampleRate: 16000,\n  },\n  {\n    id: \"whisper-base.en\",\n    repo: \"onnx-community/whisper-base.en\",\n    description: \"Base English-only model, good balance\",\n    size: \"74M\",\n    multilingual: false,\n    languages: [\"en\"],\n    sampleRate: 16000,\n  },\n  {\n    id: \"whisper-base\",\n    repo: \"onnx-community/whisper-base\",\n    description: \"Base multilingual model\",\n    size: \"74M\",\n    multilingual: true,\n    languages: [\"en\", \"es\", \"fr\", \"de\", \"it\", \"pt\", \"nl\", \"ru\", \"zh\", \"ja\", \"ko\"],\n    sampleRate: 16000,\n  },\n  {\n    id: \"whisper-small.en\",\n    repo: \"onnx-community/whisper-small.en\",\n    description: \"Small English-only model, high quality\",\n    size: \"244M\",\n    multilingual: false,\n    languages: [\"en\"],\n    sampleRate: 16000,\n  },\n  {\n    id: \"whisper-small\",\n    repo: \"onnx-community/whisper-small\",\n    description: \"Small multilingual model\",\n    size: \"244M\",\n    multilingual: true,\n    languages: [\"en\", \"es\", \"fr\", \"de\", \"it\", \"pt\", \"nl\", \"ru\", \"zh\", \"ja\", \"ko\"],\n    sampleRate: 16000,\n  },\n  {\n    id: \"whisper-large-v3-turbo\",\n    repo: \"onnx-community/whisper-large-v3-turbo\",\n    description: \"Large Turbo model, 5.4x faster, 80+ languages\",\n    size: \"809M\",\n    multilingual: true,\n    languages: [\n      \"en\",\n      \"es\",\n      \"fr\",\n      \"de\",\n      \"it\",\n      \"pt\",\n      \"nl\",\n      \"ru\",\n      \"zh\",\n      \"ja\",\n      \"ko\",\n      \"ar\",\n      \"hi\",\n      \"vi\",\n      \"th\",\n    ],\n    sampleRate: 16000,\n  },\n];\n\n// Default model\nconst DEFAULT_MODEL = \"whisper-tiny.en\";\n\n// ============================================\n// Audio Utilities\n// ============================================\n\n/**\n * Decode WAV file to Float32Array\n * Handles stereo to mono conversion\n */\nexport function decodeWav(buffer: Uint8Array): { audio: Float32Array; sampleRate: number } {\n  const view = new DataView(buffer.buffer, buffer.byteOffset, buffer.byteLength);\n\n  // Validate RIFF header\n  const riff = String.fromCharCode(buffer[0], buffer[1], buffer[2], buffer[3]);\n  if (riff !== \"RIFF\") {\n    throw new Error(\"Invalid WAV file: missing RIFF header\");\n  }\n\n  // Get format details\n  const numChannels = view.getUint16(22, true);\n  const sampleRate = view.getUint32(24, true);\n  const bitsPerSample = view.getUint16(34, true);\n\n  if (bitsPerSample !== 16) {\n    throw new Error(`Unsupported bit depth: ${bitsPerSample}. Only 16-bit WAV is supported.`);\n  }\n\n  // Find data chunk\n  let dataOffset = 12;\n  while (dataOffset < buffer.length - 8) {\n    const chunkId = String.fromCharCode(\n      buffer[dataOffset],\n      buffer[dataOffset + 1],\n      buffer[dataOffset + 2],\n      buffer[dataOffset + 3],\n    );\n    const chunkSize = view.getUint32(dataOffset + 4, true);\n    if (chunkId === \"data\") {\n      dataOffset += 8;\n      break;\n    }\n    dataOffset += 8 + chunkSize;\n  }\n\n  const dataSize = buffer.length - dataOffset;\n  const bytesPerSample = bitsPerSample / 8;\n  const totalSamples = Math.floor(dataSize / bytesPerSample);\n  const samplesPerChannel = Math.floor(totalSamples / numChannels);\n\n  // Convert to mono Float32Array\n  const audio = new Float32Array(samplesPerChannel);\n\n  for (let i = 0; i < samplesPerChannel; i++) {\n    if (numChannels === 2) {\n      const left = view.getInt16(dataOffset + i * 4, true);\n      const right = view.getInt16(dataOffset + i * 4 + 2, true);\n      audio[i] = (left + right) / 2 / 32768;\n    } else {\n      const sample = view.getInt16(dataOffset + i * 2, true);\n      audio[i] = sample / 32768;\n    }\n  }\n\n  return { audio, sampleRate };\n}\n\n/**\n * Resample audio to target sample rate using linear interpolation\n */\nexport function resampleAudio(audio: Float32Array, fromRate: number, toRate: number): Float32Array {\n  if (fromRate === toRate) return audio;\n\n  const ratio = toRate / fromRate;\n  const newLength = Math.round(audio.length * ratio);\n  const result = new Float32Array(newLength);\n\n  for (let i = 0; i < newLength; i++) {\n    const srcIndex = i / ratio;\n    const floor = Math.floor(srcIndex);\n    const ceil = Math.min(floor + 1, audio.length - 1);\n    const t = srcIndex - floor;\n    result[i] = audio[floor] * (1 - t) + audio[ceil] * t;\n  }\n\n  return result;\n}\n\n// ============================================\n// WhisperSTT Class\n// ============================================\n\n/**\n * Speech-to-Text using Whisper ONNX models\n */\nexport class WhisperSTT {\n  private modelConfig: STTModelConfig;\n  private pipeline: any = null;\n  private loadPromise: Promise<void> | null = null;\n  private _isLoaded = false;\n  private _deviceMode: \"webgpu\" | \"cpu\" = \"cpu\";\n\n  constructor(modelId: string = DEFAULT_MODEL) {\n    const config = WHISPER_MODELS.find((m) => m.id === modelId);\n    if (!config) {\n      const available = WHISPER_MODELS.map((m) => m.id).join(\", \");\n      throw new Error(`Unknown STT model: ${modelId}. Available: ${available}`);\n    }\n    this.modelConfig = config;\n  }\n\n  /**\n   * Check if model is loaded\n   */\n  isLoaded(): boolean {\n    return this._isLoaded;\n  }\n\n  /**\n   * Get model configuration\n   */\n  getModelConfig(): STTModelConfig {\n    return this.modelConfig;\n  }\n\n  /**\n   * Get model info (alias for getModelConfig)\n   */\n  getModelInfo(): STTModelConfig {\n    return this.modelConfig;\n  }\n\n  /**\n   * Get current device mode\n   */\n  getDeviceMode(): \"webgpu\" | \"cpu\" {\n    return this._deviceMode;\n  }\n\n  /**\n   * List available models\n   */\n  static listModels(): STTModelConfig[] {\n    return [...WHISPER_MODELS];\n  }\n\n  /**\n   * Load the STT model\n   */\n  async load(options: LoadSTTOptions = {}): Promise<void> {\n    if (this._isLoaded) return;\n    if (this.loadPromise) {\n      await this.loadPromise;\n      return;\n    }\n\n    this.loadPromise = this._load(options);\n    await this.loadPromise;\n  }\n\n  private async _load(options: LoadSTTOptions = {}): Promise<void> {\n    const { onProgress, device = \"auto\" } = options;\n\n    onProgress?.({ status: \"Loading transformers.js...\" });\n\n    // Check if we're in Node.js or browser\n    const isNode = typeof process !== \"undefined\" && process.versions?.node;\n\n    // Import transformers.js dynamically\n    // tsdown handles resolution: Node.js = external, Browser = bundled\n    const { pipeline, env } = await import(\"@huggingface/transformers\");\n\n    // Configure environment based on runtime\n    if (isNode) {\n      // Node.js: allow local models (for CLI/server use)\n      env.allowLocalModels = true;\n      env.allowRemoteModels = true;\n    } else {\n      // Browser: use IndexedDB cache, fetch from HuggingFace CDN\n      env.useBrowserCache = true;\n      env.allowLocalModels = false;\n    }\n\n    // Determine device\n    // Note: Whisper ONNX models work best with fp32 on CPU/WASM\n    // WebGPU support for ASR is limited, so we use CPU for reliability\n    let tfDevice: \"webgpu\" | \"cpu\" | \"wasm\" = \"cpu\";\n\n    // In browser, use WASM for better compatibility\n    if (!isNode) {\n      tfDevice = \"wasm\";\n    }\n\n    // Store device mode\n    this._deviceMode = \"cpu\"; // STT always reports as CPU since WASM is CPU-based\n\n    onProgress?.({ status: `Loading ${this.modelConfig.id}...` });\n\n    // Load the ASR pipeline\n    // Always use fp32 for Whisper models (fp16 not available for ONNX ASR)\n    this.pipeline = await pipeline(\"automatic-speech-recognition\", this.modelConfig.repo, {\n      dtype: \"fp32\",\n      device: tfDevice,\n      progress_callback: (progress: any) => {\n        if (progress.status === \"progress\" && progress.file) {\n          onProgress?.({\n            status: `Downloading ${progress.file}`,\n            progress: Math.round(progress.progress || 0),\n            file: progress.file,\n          });\n        }\n      },\n    });\n\n    this._isLoaded = true;\n    onProgress?.({ status: `Ready (${tfDevice.toUpperCase()})!` });\n  }\n\n  /**\n   * Transcribe audio to text\n   *\n   * @param audio - Audio data as Float32Array (mono, 16kHz preferred) or Uint8Array (WAV file)\n   * @param options - Transcription options\n   * @returns Transcription result with text and optional timestamps\n   */\n  async transcribe(\n    audio: Float32Array | Uint8Array,\n    options: TranscribeOptions = {},\n  ): Promise<TranscribeResult> {\n    if (!this._isLoaded) {\n      throw new Error(\"STT model not loaded. Call load() first.\");\n    }\n\n    const { language, timestamps = false, onProgress } = options;\n    const startTime = performance.now();\n\n    // Convert Uint8Array (WAV) to Float32Array\n    let audioData: Float32Array;\n    let inputSampleRate = 16000;\n\n    if (audio instanceof Uint8Array) {\n      onProgress?.({ status: \"Decoding audio...\" });\n      const decoded = decodeWav(audio);\n      audioData = decoded.audio;\n      inputSampleRate = decoded.sampleRate;\n    } else {\n      audioData = audio;\n    }\n\n    // Resample to 16kHz if needed\n    if (inputSampleRate !== 16000) {\n      onProgress?.({ status: \"Resampling to 16kHz...\" });\n      audioData = resampleAudio(audioData, inputSampleRate, 16000);\n    }\n\n    const audioDuration = audioData.length / 16000;\n    onProgress?.({ status: `Transcribing ${audioDuration.toFixed(1)}s of audio...` });\n\n    // Build pipeline options\n    const pipelineOptions: any = {};\n\n    // Only set language for multilingual models\n    if (language && this.modelConfig.multilingual) {\n      pipelineOptions.language = language;\n      pipelineOptions.task = \"transcribe\";\n    }\n\n    // Enable timestamps if requested\n    if (timestamps) {\n      pipelineOptions.return_timestamps = true;\n    }\n\n    // Run transcription\n    const result = await this.pipeline(audioData, pipelineOptions);\n\n    const totalTime = performance.now() - startTime;\n\n    // Build result\n    let text = result.text?.trim() || \"\";\n\n    // Filter out Whisper artifacts\n    if (text === \"[BLANK_AUDIO]\" || text === \"(blank audio)\" || text === \"[BLANK AUDIO]\") {\n      text = \"\";\n    }\n\n    const transcribeResult: TranscribeResult = {\n      text,\n      language: language || (this.modelConfig.multilingual ? \"auto\" : \"en\"),\n      duration: audioDuration,\n      totalTime,\n    };\n\n    // Add segments if timestamps were requested\n    if (timestamps && result.chunks) {\n      transcribeResult.segments = result.chunks.map(\n        (chunk: any): TranscribeSegment => ({\n          text: chunk.text?.trim() || \"\",\n          start: chunk.timestamp?.[0] || 0,\n          end: chunk.timestamp?.[1] || 0,\n        }),\n      );\n    }\n\n    onProgress?.({ status: \"Done!\" });\n\n    return transcribeResult;\n  }\n\n  /**\n   * Create a streaming transcription session\n   *\n   * Transcribes audio in real-time by processing chunks at regular intervals.\n   * Perfect for live captioning, call transcription, or real-time subtitles.\n   *\n   * @param options - Streaming options\n   * @returns Streaming session controller\n   *\n   * @example\n   * ```ts\n   * const session = stt.createStreamingSession({\n   *   chunkDuration: 3000,  // Transcribe every 3 seconds\n   *   onChunk: (text, idx) => console.log(`Chunk ${idx}: ${text}`),\n   *   onTranscript: (fullText) => console.log(\"Full:\", fullText),\n   * });\n   *\n   * // Feed audio data as it comes in (Float32Array at 16kHz)\n   * session.feedAudio(audioChunk);\n   *\n   * // Or manually trigger transcription\n   * await session.flush();\n   *\n   * // Stop and get final transcript\n   * const finalText = await session.stop();\n   * ```\n   */\n  createStreamingSession(\n    options: StreamingTranscriptionOptions = {},\n  ): StreamingTranscriptionSession {\n    const {\n      chunkDuration = 3000,\n      minChunkSize = 8000, // ~0.5 seconds at 16kHz\n      onChunk,\n      onTranscript,\n      onError,\n      language,\n    } = options;\n\n    let audioBuffer: Float32Array[] = [];\n    let fullTranscript = \"\";\n    let chunkIndex = 0;\n    let intervalId: ReturnType<typeof setInterval> | null = null;\n    let isRunning = false;\n\n    const getBufferSize = (): number => {\n      return audioBuffer.reduce((sum, chunk) => sum + chunk.length, 0);\n    };\n\n    const mergeBuffer = (): Float32Array => {\n      const totalLength = getBufferSize();\n      const merged = new Float32Array(totalLength);\n      let offset = 0;\n      for (const chunk of audioBuffer) {\n        merged.set(chunk, offset);\n        offset += chunk.length;\n      }\n      return merged;\n    };\n\n    const transcribeBuffer = async (): Promise<string> => {\n      if (!this._isLoaded || getBufferSize() < minChunkSize) {\n        return \"\";\n      }\n\n      const audio = mergeBuffer();\n      audioBuffer = []; // Clear buffer\n\n      try {\n        const result = await this.transcribe(audio, { language });\n        const text = result.text.trim();\n\n        if (text) {\n          chunkIndex++;\n          onChunk?.(text, chunkIndex);\n\n          // Append to full transcript\n          fullTranscript = fullTranscript + (fullTranscript ? \" \" : \"\") + text;\n          onTranscript?.(fullTranscript);\n        }\n\n        return text;\n      } catch (e: any) {\n        onError?.(e.message || \"Transcription failed\");\n        return \"\";\n      }\n    };\n\n    let aborted = false;\n\n    const session: StreamingTranscriptionSession = {\n      feedAudio: (audio: Float32Array) => {\n        if (!aborted) {\n          audioBuffer.push(audio);\n        }\n      },\n\n      flush: async () => {\n        if (aborted) return \"\";\n        return transcribeBuffer();\n      },\n\n      start: () => {\n        if (isRunning || aborted) return;\n        isRunning = true;\n\n        intervalId = setInterval(async () => {\n          if (isRunning && !aborted) {\n            await transcribeBuffer();\n          }\n        }, chunkDuration);\n      },\n\n      stop: async () => {\n        isRunning = false;\n\n        if (intervalId) {\n          clearInterval(intervalId);\n          intervalId = null;\n        }\n\n        // Transcribe any remaining audio (unless aborted)\n        if (!aborted && getBufferSize() >= minChunkSize) {\n          await transcribeBuffer();\n        }\n\n        return fullTranscript;\n      },\n\n      abort: () => {\n        // Immediately stop without final transcription\n        aborted = true;\n        isRunning = false;\n\n        if (intervalId) {\n          clearInterval(intervalId);\n          intervalId = null;\n        }\n\n        audioBuffer = [];\n      },\n\n      isRunning: () => isRunning,\n\n      getTranscript: () => fullTranscript,\n\n      getChunkCount: () => chunkIndex,\n\n      reset: () => {\n        audioBuffer = [];\n        fullTranscript = \"\";\n        chunkIndex = 0;\n      },\n    };\n\n    return session;\n  }\n\n  /**\n   * Dispose of resources\n   */\n  dispose(): void {\n    this.pipeline = null;\n    this._isLoaded = false;\n    this.loadPromise = null;\n  }\n}\n"],"mappings":";;;;;AA0CA,MAAaA,iBAAmC;CAC9C;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW,CAAC,KAAK;EACjB,YAAY;EACb;CACD;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW;GAAC;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAK;EAC7E,YAAY;EACb;CACD;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW,CAAC,KAAK;EACjB,YAAY;EACb;CACD;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW;GAAC;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAK;EAC7E,YAAY;EACb;CACD;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW,CAAC,KAAK;EACjB,YAAY;EACb;CACD;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW;GAAC;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAK;EAC7E,YAAY;EACb;CACD;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW;GACT;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACD;EACD,YAAY;EACb;CACF;AAGD,MAAM,gBAAgB;;;;;AAUtB,SAAgB,UAAU,QAAiE;CACzF,MAAM,OAAO,IAAI,SAAS,OAAO,QAAQ,OAAO,YAAY,OAAO,WAAW;AAI9E,KADa,OAAO,aAAa,OAAO,IAAI,OAAO,IAAI,OAAO,IAAI,OAAO,GAAG,KAC/D,OACX,OAAM,IAAI,MAAM,wCAAwC;CAI1D,MAAM,cAAc,KAAK,UAAU,IAAI,KAAK;CAC5C,MAAM,aAAa,KAAK,UAAU,IAAI,KAAK;CAC3C,MAAM,gBAAgB,KAAK,UAAU,IAAI,KAAK;AAE9C,KAAI,kBAAkB,GACpB,OAAM,IAAI,MAAM,0BAA0B,cAAc,iCAAiC;CAI3F,IAAI,aAAa;AACjB,QAAO,aAAa,OAAO,SAAS,GAAG;EACrC,MAAM,UAAU,OAAO,aACrB,OAAO,aACP,OAAO,aAAa,IACpB,OAAO,aAAa,IACpB,OAAO,aAAa,GACrB;EACD,MAAM,YAAY,KAAK,UAAU,aAAa,GAAG,KAAK;AACtD,MAAI,YAAY,QAAQ;AACtB,iBAAc;AACd;;AAEF,gBAAc,IAAI;;CAGpB,MAAM,WAAW,OAAO,SAAS;CACjC,MAAM,iBAAiB,gBAAgB;CACvC,MAAM,eAAe,KAAK,MAAM,WAAW,eAAe;CAC1D,MAAM,oBAAoB,KAAK,MAAM,eAAe,YAAY;CAGhE,MAAM,QAAQ,IAAI,aAAa,kBAAkB;AAEjD,MAAK,IAAI,IAAI,GAAG,IAAI,mBAAmB,IACrC,KAAI,gBAAgB,EAGlB,OAAM,MAFO,KAAK,SAAS,aAAa,IAAI,GAAG,KAAK,GACtC,KAAK,SAAS,aAAa,IAAI,IAAI,GAAG,KAAK,IAC7B,IAAI;KAGhC,OAAM,KADS,KAAK,SAAS,aAAa,IAAI,GAAG,KAAK,GAClC;AAIxB,QAAO;EAAE;EAAO;EAAY;;;;;AAM9B,SAAgB,cAAc,OAAqB,UAAkB,QAA8B;AACjG,KAAI,aAAa,OAAQ,QAAO;CAEhC,MAAM,QAAQ,SAAS;CACvB,MAAM,YAAY,KAAK,MAAM,MAAM,SAAS,MAAM;CAClD,MAAM,SAAS,IAAI,aAAa,UAAU;AAE1C,MAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;EAClC,MAAM,WAAW,IAAI;EACrB,MAAM,QAAQ,KAAK,MAAM,SAAS;EAClC,MAAM,OAAO,KAAK,IAAI,QAAQ,GAAG,MAAM,SAAS,EAAE;EAClD,MAAM,IAAI,WAAW;AACrB,SAAO,KAAK,MAAM,UAAU,IAAI,KAAK,MAAM,QAAQ;;AAGrD,QAAO;;;;;AAUT,IAAa,aAAb,MAAwB;CACtB,AAAQ;CACR,AAAQ,WAAgB;CACxB,AAAQ,cAAoC;CAC5C,AAAQ,YAAY;CACpB,AAAQ,cAAgC;CAExC,YAAY,UAAkB,eAAe;EAC3C,MAAM,SAAS,eAAe,MAAM,MAAM,EAAE,OAAO,QAAQ;AAC3D,MAAI,CAAC,QAAQ;GACX,MAAM,YAAY,eAAe,KAAK,MAAM,EAAE,GAAG,CAAC,KAAK,KAAK;AAC5D,SAAM,IAAI,MAAM,sBAAsB,QAAQ,eAAe,YAAY;;AAE3E,OAAK,cAAc;;;;;CAMrB,WAAoB;AAClB,SAAO,KAAK;;;;;CAMd,iBAAiC;AAC/B,SAAO,KAAK;;;;;CAMd,eAA+B;AAC7B,SAAO,KAAK;;;;;CAMd,gBAAkC;AAChC,SAAO,KAAK;;;;;CAMd,OAAO,aAA+B;AACpC,SAAO,CAAC,GAAG,eAAe;;;;;CAM5B,MAAM,KAAK,UAA0B,EAAE,EAAiB;AACtD,MAAI,KAAK,UAAW;AACpB,MAAI,KAAK,aAAa;AACpB,SAAM,KAAK;AACX;;AAGF,OAAK,cAAc,KAAK,MAAM,QAAQ;AACtC,QAAM,KAAK;;CAGb,MAAc,MAAM,UAA0B,EAAE,EAAiB;EAC/D,MAAM,EAAE,YAAY,SAAS,WAAW;AAExC,eAAa,EAAE,QAAQ,8BAA8B,CAAC;EAGtD,MAAM,SAAS,OAAO,YAAY,eAAe,QAAQ,UAAU;EAInE,MAAM,EAAE,UAAU,QAAQ,MAAM,OAAO;AAGvC,MAAI,QAAQ;AAEV,OAAI,mBAAmB;AACvB,OAAI,oBAAoB;SACnB;AAEL,OAAI,kBAAkB;AACtB,OAAI,mBAAmB;;EAMzB,IAAIC,WAAsC;AAG1C,MAAI,CAAC,OACH,YAAW;AAIb,OAAK,cAAc;AAEnB,eAAa,EAAE,QAAQ,WAAW,KAAK,YAAY,GAAG,MAAM,CAAC;AAI7D,OAAK,WAAW,MAAM,SAAS,gCAAgC,KAAK,YAAY,MAAM;GACpF,OAAO;GACP,QAAQ;GACR,oBAAoB,aAAkB;AACpC,QAAI,SAAS,WAAW,cAAc,SAAS,KAC7C,cAAa;KACX,QAAQ,eAAe,SAAS;KAChC,UAAU,KAAK,MAAM,SAAS,YAAY,EAAE;KAC5C,MAAM,SAAS;KAChB,CAAC;;GAGP,CAAC;AAEF,OAAK,YAAY;AACjB,eAAa,EAAE,QAAQ,UAAU,SAAS,aAAa,CAAC,KAAK,CAAC;;;;;;;;;CAUhE,MAAM,WACJ,OACA,UAA6B,EAAE,EACJ;AAC3B,MAAI,CAAC,KAAK,UACR,OAAM,IAAI,MAAM,2CAA2C;EAG7D,MAAM,EAAE,UAAU,aAAa,OAAO,eAAe;EACrD,MAAM,YAAY,YAAY,KAAK;EAGnC,IAAIC;EACJ,IAAI,kBAAkB;AAEtB,MAAI,iBAAiB,YAAY;AAC/B,gBAAa,EAAE,QAAQ,qBAAqB,CAAC;GAC7C,MAAM,UAAU,UAAU,MAAM;AAChC,eAAY,QAAQ;AACpB,qBAAkB,QAAQ;QAE1B,aAAY;AAId,MAAI,oBAAoB,MAAO;AAC7B,gBAAa,EAAE,QAAQ,0BAA0B,CAAC;AAClD,eAAY,cAAc,WAAW,iBAAiB,KAAM;;EAG9D,MAAM,gBAAgB,UAAU,SAAS;AACzC,eAAa,EAAE,QAAQ,gBAAgB,cAAc,QAAQ,EAAE,CAAC,gBAAgB,CAAC;EAGjF,MAAMC,kBAAuB,EAAE;AAG/B,MAAI,YAAY,KAAK,YAAY,cAAc;AAC7C,mBAAgB,WAAW;AAC3B,mBAAgB,OAAO;;AAIzB,MAAI,WACF,iBAAgB,oBAAoB;EAItC,MAAM,SAAS,MAAM,KAAK,SAAS,WAAW,gBAAgB;EAE9D,MAAM,YAAY,YAAY,KAAK,GAAG;EAGtC,IAAI,OAAO,OAAO,MAAM,MAAM,IAAI;AAGlC,MAAI,SAAS,mBAAmB,SAAS,mBAAmB,SAAS,gBACnE,QAAO;EAGT,MAAMC,mBAAqC;GACzC;GACA,UAAU,aAAa,KAAK,YAAY,eAAe,SAAS;GAChE,UAAU;GACV;GACD;AAGD,MAAI,cAAc,OAAO,OACvB,kBAAiB,WAAW,OAAO,OAAO,KACvC,WAAmC;GAClC,MAAM,MAAM,MAAM,MAAM,IAAI;GAC5B,OAAO,MAAM,YAAY,MAAM;GAC/B,KAAK,MAAM,YAAY,MAAM;GAC9B,EACF;AAGH,eAAa,EAAE,QAAQ,SAAS,CAAC;AAEjC,SAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8BT,uBACE,UAAyC,EAAE,EACZ;EAC/B,MAAM,EACJ,gBAAgB,KAChB,eAAe,KACf,SACA,cACA,SACA,aACE;EAEJ,IAAIC,cAA8B,EAAE;EACpC,IAAI,iBAAiB;EACrB,IAAI,aAAa;EACjB,IAAIC,aAAoD;EACxD,IAAI,YAAY;EAEhB,MAAM,sBAA8B;AAClC,UAAO,YAAY,QAAQ,KAAK,UAAU,MAAM,MAAM,QAAQ,EAAE;;EAGlE,MAAM,oBAAkC;GACtC,MAAM,cAAc,eAAe;GACnC,MAAM,SAAS,IAAI,aAAa,YAAY;GAC5C,IAAI,SAAS;AACb,QAAK,MAAM,SAAS,aAAa;AAC/B,WAAO,IAAI,OAAO,OAAO;AACzB,cAAU,MAAM;;AAElB,UAAO;;EAGT,MAAM,mBAAmB,YAA6B;AACpD,OAAI,CAAC,KAAK,aAAa,eAAe,GAAG,aACvC,QAAO;GAGT,MAAM,QAAQ,aAAa;AAC3B,iBAAc,EAAE;AAEhB,OAAI;IAEF,MAAM,QADS,MAAM,KAAK,WAAW,OAAO,EAAE,UAAU,CAAC,EACrC,KAAK,MAAM;AAE/B,QAAI,MAAM;AACR;AACA,eAAU,MAAM,WAAW;AAG3B,sBAAiB,kBAAkB,iBAAiB,MAAM,MAAM;AAChE,oBAAe,eAAe;;AAGhC,WAAO;YACAC,GAAQ;AACf,cAAU,EAAE,WAAW,uBAAuB;AAC9C,WAAO;;;EAIX,IAAI,UAAU;AAmEd,SAjE+C;GAC7C,YAAY,UAAwB;AAClC,QAAI,CAAC,QACH,aAAY,KAAK,MAAM;;GAI3B,OAAO,YAAY;AACjB,QAAI,QAAS,QAAO;AACpB,WAAO,kBAAkB;;GAG3B,aAAa;AACX,QAAI,aAAa,QAAS;AAC1B,gBAAY;AAEZ,iBAAa,YAAY,YAAY;AACnC,SAAI,aAAa,CAAC,QAChB,OAAM,kBAAkB;OAEzB,cAAc;;GAGnB,MAAM,YAAY;AAChB,gBAAY;AAEZ,QAAI,YAAY;AACd,mBAAc,WAAW;AACzB,kBAAa;;AAIf,QAAI,CAAC,WAAW,eAAe,IAAI,aACjC,OAAM,kBAAkB;AAG1B,WAAO;;GAGT,aAAa;AAEX,cAAU;AACV,gBAAY;AAEZ,QAAI,YAAY;AACd,mBAAc,WAAW;AACzB,kBAAa;;AAGf,kBAAc,EAAE;;GAGlB,iBAAiB;GAEjB,qBAAqB;GAErB,qBAAqB;GAErB,aAAa;AACX,kBAAc,EAAE;AAChB,qBAAiB;AACjB,iBAAa;;GAEhB;;;;;CAQH,UAAgB;AACd,OAAK,WAAW;AAChB,OAAK,YAAY;AACjB,OAAK,cAAc"}
+{"version":3,"file":"stt-Bu-E23Sc.js","names":["WHISPER_MODELS: STTModelConfig[]","tfDevice: \"webgpu\" | \"cpu\" | \"wasm\"","audioData: Float32Array","pipelineOptions: any","transcribeResult: TranscribeResult","audioBuffer: Float32Array[]","intervalId: ReturnType<typeof setInterval> | null","e: any"],"sources":["../src/core/stt.ts"],"sourcesContent":["/**\n * Speech-to-Text with Whisper\n *\n * Provides local speech recognition using Whisper ONNX models via transformers.js.\n * Supports multiple model sizes and languages.\n *\n * @example\n * ```ts\n * const stt = new WhisperSTT();\n * await stt.load({ onProgress: (p) => console.log(p.status) });\n *\n * // Transcribe audio (Float32Array at 16kHz)\n * const result = await stt.transcribe(audioData);\n * console.log(result.text);\n *\n * // With timestamps\n * const result = await stt.transcribe(audioData, { timestamps: true });\n * for (const seg of result.segments) {\n *   console.log(`[${seg.start.toFixed(1)}s] ${seg.text}`);\n * }\n * ```\n */\n\nimport type {\n  LoadSTTOptions,\n  ProgressInfo,\n  STTModelConfig,\n  StreamingTranscriptionOptions,\n  StreamingTranscriptionSession,\n  TranscribeOptions,\n  TranscribeResult,\n  TranscribeSegment,\n} from \"./types.js\";\n\n// ============================================\n// Model Registry\n// ============================================\n\n/**\n * Available Whisper models\n * Ordered by size (smallest first)\n */\nexport const WHISPER_MODELS: STTModelConfig[] = [\n  {\n    id: \"whisper-tiny.en\",\n    repo: \"onnx-community/whisper-tiny.en\",\n    description: \"Tiny English-only model, fastest\",\n    size: \"39M\",\n    multilingual: false,\n    languages: [\"en\"],\n    sampleRate: 16000,\n  },\n  {\n    id: \"whisper-tiny\",\n    repo: \"onnx-community/whisper-tiny\",\n    description: \"Tiny multilingual model\",\n    size: \"39M\",\n    multilingual: true,\n    languages: [\"en\", \"es\", \"fr\", \"de\", \"it\", \"pt\", \"nl\", \"ru\", \"zh\", \"ja\", \"ko\"],\n    sampleRate: 16000,\n  },\n  {\n    id: \"whisper-base.en\",\n    repo: \"onnx-community/whisper-base.en\",\n    description: \"Base English-only model, good balance\",\n    size: \"74M\",\n    multilingual: false,\n    languages: [\"en\"],\n    sampleRate: 16000,\n  },\n  {\n    id: \"whisper-base\",\n    repo: \"onnx-community/whisper-base\",\n    description: \"Base multilingual model\",\n    size: \"74M\",\n    multilingual: true,\n    languages: [\"en\", \"es\", \"fr\", \"de\", \"it\", \"pt\", \"nl\", \"ru\", \"zh\", \"ja\", \"ko\"],\n    sampleRate: 16000,\n  },\n  {\n    id: \"whisper-small.en\",\n    repo: \"onnx-community/whisper-small.en\",\n    description: \"Small English-only model, high quality\",\n    size: \"244M\",\n    multilingual: false,\n    languages: [\"en\"],\n    sampleRate: 16000,\n  },\n  {\n    id: \"whisper-small\",\n    repo: \"onnx-community/whisper-small\",\n    description: \"Small multilingual model\",\n    size: \"244M\",\n    multilingual: true,\n    languages: [\"en\", \"es\", \"fr\", \"de\", \"it\", \"pt\", \"nl\", \"ru\", \"zh\", \"ja\", \"ko\"],\n    sampleRate: 16000,\n  },\n  {\n    id: \"whisper-large-v3-turbo\",\n    repo: \"onnx-community/whisper-large-v3-turbo\",\n    description: \"Large Turbo model, 5.4x faster, 80+ languages\",\n    size: \"809M\",\n    multilingual: true,\n    languages: [\n      \"en\",\n      \"es\",\n      \"fr\",\n      \"de\",\n      \"it\",\n      \"pt\",\n      \"nl\",\n      \"ru\",\n      \"zh\",\n      \"ja\",\n      \"ko\",\n      \"ar\",\n      \"hi\",\n      \"vi\",\n      \"th\",\n    ],\n    sampleRate: 16000,\n  },\n];\n\n// Default model\nconst DEFAULT_MODEL = \"whisper-tiny.en\";\n\n// ============================================\n// Audio Utilities\n// ============================================\n\n/**\n * Decode WAV file to Float32Array\n * Handles stereo to mono conversion\n */\nexport function decodeWav(buffer: Uint8Array): { audio: Float32Array; sampleRate: number } {\n  const view = new DataView(buffer.buffer, buffer.byteOffset, buffer.byteLength);\n\n  // Validate RIFF header\n  const riff = String.fromCharCode(buffer[0], buffer[1], buffer[2], buffer[3]);\n  if (riff !== \"RIFF\") {\n    throw new Error(\"Invalid WAV file: missing RIFF header\");\n  }\n\n  // Get format details\n  const numChannels = view.getUint16(22, true);\n  const sampleRate = view.getUint32(24, true);\n  const bitsPerSample = view.getUint16(34, true);\n\n  if (bitsPerSample !== 16) {\n    throw new Error(`Unsupported bit depth: ${bitsPerSample}. Only 16-bit WAV is supported.`);\n  }\n\n  // Find data chunk\n  let dataOffset = 12;\n  while (dataOffset < buffer.length - 8) {\n    const chunkId = String.fromCharCode(\n      buffer[dataOffset],\n      buffer[dataOffset + 1],\n      buffer[dataOffset + 2],\n      buffer[dataOffset + 3],\n    );\n    const chunkSize = view.getUint32(dataOffset + 4, true);\n    if (chunkId === \"data\") {\n      dataOffset += 8;\n      break;\n    }\n    dataOffset += 8 + chunkSize;\n  }\n\n  const dataSize = buffer.length - dataOffset;\n  const bytesPerSample = bitsPerSample / 8;\n  const totalSamples = Math.floor(dataSize / bytesPerSample);\n  const samplesPerChannel = Math.floor(totalSamples / numChannels);\n\n  // Convert to mono Float32Array\n  const audio = new Float32Array(samplesPerChannel);\n\n  for (let i = 0; i < samplesPerChannel; i++) {\n    if (numChannels === 2) {\n      const left = view.getInt16(dataOffset + i * 4, true);\n      const right = view.getInt16(dataOffset + i * 4 + 2, true);\n      audio[i] = (left + right) / 2 / 32768;\n    } else {\n      const sample = view.getInt16(dataOffset + i * 2, true);\n      audio[i] = sample / 32768;\n    }\n  }\n\n  return { audio, sampleRate };\n}\n\n/**\n * Resample audio to target sample rate using linear interpolation\n */\nexport function resampleAudio(audio: Float32Array, fromRate: number, toRate: number): Float32Array {\n  if (fromRate === toRate) return audio;\n\n  const ratio = toRate / fromRate;\n  const newLength = Math.round(audio.length * ratio);\n  const result = new Float32Array(newLength);\n\n  for (let i = 0; i < newLength; i++) {\n    const srcIndex = i / ratio;\n    const floor = Math.floor(srcIndex);\n    const ceil = Math.min(floor + 1, audio.length - 1);\n    const t = srcIndex - floor;\n    result[i] = audio[floor] * (1 - t) + audio[ceil] * t;\n  }\n\n  return result;\n}\n\n// ============================================\n// WhisperSTT Class\n// ============================================\n\n/**\n * Speech-to-Text using Whisper ONNX models\n */\nexport class WhisperSTT {\n  private modelConfig: STTModelConfig;\n  private pipeline: any = null;\n  private loadPromise: Promise<void> | null = null;\n  private _isLoaded = false;\n  private _deviceMode: \"webgpu\" | \"cpu\" = \"cpu\";\n\n  constructor(modelId: string = DEFAULT_MODEL) {\n    const config = WHISPER_MODELS.find((m) => m.id === modelId);\n    if (!config) {\n      const available = WHISPER_MODELS.map((m) => m.id).join(\", \");\n      throw new Error(`Unknown STT model: ${modelId}. Available: ${available}`);\n    }\n    this.modelConfig = config;\n  }\n\n  /**\n   * Check if model is loaded\n   */\n  isLoaded(): boolean {\n    return this._isLoaded;\n  }\n\n  /**\n   * Get model configuration\n   */\n  getModelConfig(): STTModelConfig {\n    return this.modelConfig;\n  }\n\n  /**\n   * Get model info (alias for getModelConfig)\n   */\n  getModelInfo(): STTModelConfig {\n    return this.modelConfig;\n  }\n\n  /**\n   * Get current device mode\n   */\n  getDeviceMode(): \"webgpu\" | \"cpu\" {\n    return this._deviceMode;\n  }\n\n  /**\n   * List available models\n   */\n  static listModels(): STTModelConfig[] {\n    return [...WHISPER_MODELS];\n  }\n\n  /**\n   * Load the STT model\n   */\n  async load(options: LoadSTTOptions = {}): Promise<void> {\n    if (this._isLoaded) return;\n    if (this.loadPromise) {\n      await this.loadPromise;\n      return;\n    }\n\n    this.loadPromise = this._load(options);\n    await this.loadPromise;\n  }\n\n  private async _load(options: LoadSTTOptions = {}): Promise<void> {\n    const { onProgress, device = \"auto\" } = options;\n\n    onProgress?.({ status: \"Loading transformers.js...\" });\n\n    // Check if we're in Node.js or browser\n    const isNode = typeof process !== \"undefined\" && process.versions?.node;\n\n    // Import transformers.js dynamically\n    // tsdown handles resolution: Node.js = external, Browser = bundled\n    const { pipeline, env } = await import(\"@huggingface/transformers\");\n\n    // Configure environment based on runtime\n    if (isNode) {\n      // Node.js: allow local models (for CLI/server use)\n      env.allowLocalModels = true;\n      env.allowRemoteModels = true;\n    } else {\n      // Browser: use IndexedDB cache, fetch from HuggingFace CDN\n      env.useBrowserCache = true;\n      env.allowLocalModels = false;\n    }\n\n    // Determine device\n    // Note: Whisper ONNX models work best with fp32 on CPU/WASM\n    // WebGPU support for ASR is limited, so we use CPU for reliability\n    let tfDevice: \"webgpu\" | \"cpu\" | \"wasm\" = \"cpu\";\n\n    // In browser, use WASM for better compatibility\n    if (!isNode) {\n      tfDevice = \"wasm\";\n    }\n\n    // Store device mode\n    this._deviceMode = \"cpu\"; // STT always reports as CPU since WASM is CPU-based\n\n    onProgress?.({ status: `Loading ${this.modelConfig.id}...` });\n\n    // Load the ASR pipeline\n    // Always use fp32 for Whisper models (fp16 not available for ONNX ASR)\n    this.pipeline = await pipeline(\"automatic-speech-recognition\", this.modelConfig.repo, {\n      dtype: \"fp32\",\n      device: tfDevice,\n      progress_callback: (progress: any) => {\n        if (progress.status === \"progress\" && progress.file) {\n          onProgress?.({\n            status: `Downloading ${progress.file}`,\n            progress: Math.round(progress.progress || 0),\n            file: progress.file,\n          });\n        }\n      },\n    });\n\n    this._isLoaded = true;\n    onProgress?.({ status: `Ready (${tfDevice.toUpperCase()})!` });\n  }\n\n  /**\n   * Transcribe audio to text\n   *\n   * @param audio - Audio data as Float32Array (mono, 16kHz preferred) or Uint8Array (WAV file)\n   * @param options - Transcription options\n   * @returns Transcription result with text and optional timestamps\n   */\n  async transcribe(\n    audio: Float32Array | Uint8Array,\n    options: TranscribeOptions = {},\n  ): Promise<TranscribeResult> {\n    if (!this._isLoaded) {\n      throw new Error(\"STT model not loaded. Call load() first.\");\n    }\n\n    const { language, timestamps = false, onProgress } = options;\n    const startTime = performance.now();\n\n    // Convert Uint8Array (WAV) to Float32Array\n    let audioData: Float32Array;\n    let inputSampleRate = 16000;\n\n    if (audio instanceof Uint8Array) {\n      onProgress?.({ status: \"Decoding audio...\" });\n      const decoded = decodeWav(audio);\n      audioData = decoded.audio;\n      inputSampleRate = decoded.sampleRate;\n    } else {\n      audioData = audio;\n    }\n\n    // Resample to 16kHz if needed\n    if (inputSampleRate !== 16000) {\n      onProgress?.({ status: \"Resampling to 16kHz...\" });\n      audioData = resampleAudio(audioData, inputSampleRate, 16000);\n    }\n\n    const audioDuration = audioData.length / 16000;\n    onProgress?.({ status: `Transcribing ${audioDuration.toFixed(1)}s of audio...` });\n\n    // Build pipeline options\n    const pipelineOptions: any = {};\n\n    // Only set language for multilingual models\n    if (language && this.modelConfig.multilingual) {\n      pipelineOptions.language = language;\n      pipelineOptions.task = \"transcribe\";\n    }\n\n    // Enable timestamps if requested\n    if (timestamps) {\n      pipelineOptions.return_timestamps = true;\n    }\n\n    // Run transcription\n    const result = await this.pipeline(audioData, pipelineOptions);\n\n    const totalTime = performance.now() - startTime;\n\n    // Build result\n    let text = result.text?.trim() || \"\";\n\n    // Filter out Whisper artifacts\n    if (text === \"[BLANK_AUDIO]\" || text === \"(blank audio)\" || text === \"[BLANK AUDIO]\") {\n      text = \"\";\n    }\n\n    const transcribeResult: TranscribeResult = {\n      text,\n      language: language || (this.modelConfig.multilingual ? \"auto\" : \"en\"),\n      duration: audioDuration,\n      totalTime,\n    };\n\n    // Add segments if timestamps were requested\n    if (timestamps && result.chunks) {\n      transcribeResult.segments = result.chunks.map(\n        (chunk: any): TranscribeSegment => ({\n          text: chunk.text?.trim() || \"\",\n          start: chunk.timestamp?.[0] || 0,\n          end: chunk.timestamp?.[1] || 0,\n        }),\n      );\n    }\n\n    onProgress?.({ status: \"Done!\" });\n\n    return transcribeResult;\n  }\n\n  /**\n   * Create a streaming transcription session\n   *\n   * Transcribes audio in real-time by processing chunks at regular intervals.\n   * Perfect for live captioning, call transcription, or real-time subtitles.\n   *\n   * @param options - Streaming options\n   * @returns Streaming session controller\n   *\n   * @example\n   * ```ts\n   * const session = stt.createStreamingSession({\n   *   chunkDuration: 3000,  // Transcribe every 3 seconds\n   *   onChunk: (text, idx) => console.log(`Chunk ${idx}: ${text}`),\n   *   onTranscript: (fullText) => console.log(\"Full:\", fullText),\n   * });\n   *\n   * // Feed audio data as it comes in (Float32Array at 16kHz)\n   * session.feedAudio(audioChunk);\n   *\n   * // Or manually trigger transcription\n   * await session.flush();\n   *\n   * // Stop and get final transcript\n   * const finalText = await session.stop();\n   * ```\n   */\n  createStreamingSession(\n    options: StreamingTranscriptionOptions = {},\n  ): StreamingTranscriptionSession {\n    const {\n      chunkDuration = 3000,\n      minChunkSize = 8000, // ~0.5 seconds at 16kHz\n      onChunk,\n      onTranscript,\n      onError,\n      language,\n    } = options;\n\n    let audioBuffer: Float32Array[] = [];\n    let fullTranscript = \"\";\n    let chunkIndex = 0;\n    let intervalId: ReturnType<typeof setInterval> | null = null;\n    let isRunning = false;\n\n    const getBufferSize = (): number => {\n      return audioBuffer.reduce((sum, chunk) => sum + chunk.length, 0);\n    };\n\n    const mergeBuffer = (): Float32Array => {\n      const totalLength = getBufferSize();\n      const merged = new Float32Array(totalLength);\n      let offset = 0;\n      for (const chunk of audioBuffer) {\n        merged.set(chunk, offset);\n        offset += chunk.length;\n      }\n      return merged;\n    };\n\n    const transcribeBuffer = async (): Promise<string> => {\n      if (!this._isLoaded || getBufferSize() < minChunkSize) {\n        return \"\";\n      }\n\n      const audio = mergeBuffer();\n      audioBuffer = []; // Clear buffer\n\n      try {\n        const result = await this.transcribe(audio, { language });\n        const text = result.text.trim();\n\n        if (text) {\n          chunkIndex++;\n          onChunk?.(text, chunkIndex);\n\n          // Append to full transcript\n          fullTranscript = fullTranscript + (fullTranscript ? \" \" : \"\") + text;\n          onTranscript?.(fullTranscript);\n        }\n\n        return text;\n      } catch (e: any) {\n        onError?.(e.message || \"Transcription failed\");\n        return \"\";\n      }\n    };\n\n    let aborted = false;\n\n    const session: StreamingTranscriptionSession = {\n      feedAudio: (audio: Float32Array) => {\n        if (!aborted) {\n          audioBuffer.push(audio);\n        }\n      },\n\n      flush: async () => {\n        if (aborted) return \"\";\n        return transcribeBuffer();\n      },\n\n      start: () => {\n        if (isRunning || aborted) return;\n        isRunning = true;\n\n        intervalId = setInterval(async () => {\n          if (isRunning && !aborted) {\n            await transcribeBuffer();\n          }\n        }, chunkDuration);\n      },\n\n      stop: async () => {\n        isRunning = false;\n\n        if (intervalId) {\n          clearInterval(intervalId);\n          intervalId = null;\n        }\n\n        // Transcribe any remaining audio (unless aborted)\n        if (!aborted && getBufferSize() >= minChunkSize) {\n          await transcribeBuffer();\n        }\n\n        return fullTranscript;\n      },\n\n      abort: () => {\n        // Immediately stop without final transcription\n        aborted = true;\n        isRunning = false;\n\n        if (intervalId) {\n          clearInterval(intervalId);\n          intervalId = null;\n        }\n\n        audioBuffer = [];\n      },\n\n      isRunning: () => isRunning,\n\n      getTranscript: () => fullTranscript,\n\n      getChunkCount: () => chunkIndex,\n\n      reset: () => {\n        audioBuffer = [];\n        fullTranscript = \"\";\n        chunkIndex = 0;\n      },\n    };\n\n    return session;\n  }\n\n  /**\n   * Dispose of resources\n   */\n  dispose(): void {\n    this.pipeline = null;\n    this._isLoaded = false;\n    this.loadPromise = null;\n  }\n}\n"],"mappings":";;;;;AA0CA,MAAaA,iBAAmC;CAC9C;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW,CAAC,KAAK;EACjB,YAAY;EACb;CACD;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW;GAAC;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAK;EAC7E,YAAY;EACb;CACD;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW,CAAC,KAAK;EACjB,YAAY;EACb;CACD;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW;GAAC;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAK;EAC7E,YAAY;EACb;CACD;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW,CAAC,KAAK;EACjB,YAAY;EACb;CACD;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW;GAAC;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAM;GAAK;EAC7E,YAAY;EACb;CACD;EACE,IAAI;EACJ,MAAM;EACN,aAAa;EACb,MAAM;EACN,cAAc;EACd,WAAW;GACT;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACA;GACD;EACD,YAAY;EACb;CACF;AAGD,MAAM,gBAAgB;;;;;AAUtB,SAAgB,UAAU,QAAiE;CACzF,MAAM,OAAO,IAAI,SAAS,OAAO,QAAQ,OAAO,YAAY,OAAO,WAAW;AAI9E,KADa,OAAO,aAAa,OAAO,IAAI,OAAO,IAAI,OAAO,IAAI,OAAO,GAAG,KAC/D,OACX,OAAM,IAAI,MAAM,wCAAwC;CAI1D,MAAM,cAAc,KAAK,UAAU,IAAI,KAAK;CAC5C,MAAM,aAAa,KAAK,UAAU,IAAI,KAAK;CAC3C,MAAM,gBAAgB,KAAK,UAAU,IAAI,KAAK;AAE9C,KAAI,kBAAkB,GACpB,OAAM,IAAI,MAAM,0BAA0B,cAAc,iCAAiC;CAI3F,IAAI,aAAa;AACjB,QAAO,aAAa,OAAO,SAAS,GAAG;EACrC,MAAM,UAAU,OAAO,aACrB,OAAO,aACP,OAAO,aAAa,IACpB,OAAO,aAAa,IACpB,OAAO,aAAa,GACrB;EACD,MAAM,YAAY,KAAK,UAAU,aAAa,GAAG,KAAK;AACtD,MAAI,YAAY,QAAQ;AACtB,iBAAc;AACd;;AAEF,gBAAc,IAAI;;CAGpB,MAAM,WAAW,OAAO,SAAS;CACjC,MAAM,iBAAiB,gBAAgB;CACvC,MAAM,eAAe,KAAK,MAAM,WAAW,eAAe;CAC1D,MAAM,oBAAoB,KAAK,MAAM,eAAe,YAAY;CAGhE,MAAM,QAAQ,IAAI,aAAa,kBAAkB;AAEjD,MAAK,IAAI,IAAI,GAAG,IAAI,mBAAmB,IACrC,KAAI,gBAAgB,EAGlB,OAAM,MAFO,KAAK,SAAS,aAAa,IAAI,GAAG,KAAK,GACtC,KAAK,SAAS,aAAa,IAAI,IAAI,GAAG,KAAK,IAC7B,IAAI;KAGhC,OAAM,KADS,KAAK,SAAS,aAAa,IAAI,GAAG,KAAK,GAClC;AAIxB,QAAO;EAAE;EAAO;EAAY;;;;;AAM9B,SAAgB,cAAc,OAAqB,UAAkB,QAA8B;AACjG,KAAI,aAAa,OAAQ,QAAO;CAEhC,MAAM,QAAQ,SAAS;CACvB,MAAM,YAAY,KAAK,MAAM,MAAM,SAAS,MAAM;CAClD,MAAM,SAAS,IAAI,aAAa,UAAU;AAE1C,MAAK,IAAI,IAAI,GAAG,IAAI,WAAW,KAAK;EAClC,MAAM,WAAW,IAAI;EACrB,MAAM,QAAQ,KAAK,MAAM,SAAS;EAClC,MAAM,OAAO,KAAK,IAAI,QAAQ,GAAG,MAAM,SAAS,EAAE;EAClD,MAAM,IAAI,WAAW;AACrB,SAAO,KAAK,MAAM,UAAU,IAAI,KAAK,MAAM,QAAQ;;AAGrD,QAAO;;;;;AAUT,IAAa,aAAb,MAAwB;CACtB,AAAQ;CACR,AAAQ,WAAgB;CACxB,AAAQ,cAAoC;CAC5C,AAAQ,YAAY;CACpB,AAAQ,cAAgC;CAExC,YAAY,UAAkB,eAAe;EAC3C,MAAM,SAAS,eAAe,MAAM,MAAM,EAAE,OAAO,QAAQ;AAC3D,MAAI,CAAC,QAAQ;GACX,MAAM,YAAY,eAAe,KAAK,MAAM,EAAE,GAAG,CAAC,KAAK,KAAK;AAC5D,SAAM,IAAI,MAAM,sBAAsB,QAAQ,eAAe,YAAY;;AAE3E,OAAK,cAAc;;;;;CAMrB,WAAoB;AAClB,SAAO,KAAK;;;;;CAMd,iBAAiC;AAC/B,SAAO,KAAK;;;;;CAMd,eAA+B;AAC7B,SAAO,KAAK;;;;;CAMd,gBAAkC;AAChC,SAAO,KAAK;;;;;CAMd,OAAO,aAA+B;AACpC,SAAO,CAAC,GAAG,eAAe;;;;;CAM5B,MAAM,KAAK,UAA0B,EAAE,EAAiB;AACtD,MAAI,KAAK,UAAW;AACpB,MAAI,KAAK,aAAa;AACpB,SAAM,KAAK;AACX;;AAGF,OAAK,cAAc,KAAK,MAAM,QAAQ;AACtC,QAAM,KAAK;;CAGb,MAAc,MAAM,UAA0B,EAAE,EAAiB;EAC/D,MAAM,EAAE,YAAY,SAAS,WAAW;AAExC,eAAa,EAAE,QAAQ,8BAA8B,CAAC;EAGtD,MAAM,SAAS,OAAO,YAAY,eAAe,QAAQ,UAAU;EAInE,MAAM,EAAE,UAAU,QAAQ,MAAM,OAAO;AAGvC,MAAI,QAAQ;AAEV,OAAI,mBAAmB;AACvB,OAAI,oBAAoB;SACnB;AAEL,OAAI,kBAAkB;AACtB,OAAI,mBAAmB;;EAMzB,IAAIC,WAAsC;AAG1C,MAAI,CAAC,OACH,YAAW;AAIb,OAAK,cAAc;AAEnB,eAAa,EAAE,QAAQ,WAAW,KAAK,YAAY,GAAG,MAAM,CAAC;AAI7D,OAAK,WAAW,MAAM,SAAS,gCAAgC,KAAK,YAAY,MAAM;GACpF,OAAO;GACP,QAAQ;GACR,oBAAoB,aAAkB;AACpC,QAAI,SAAS,WAAW,cAAc,SAAS,KAC7C,cAAa;KACX,QAAQ,eAAe,SAAS;KAChC,UAAU,KAAK,MAAM,SAAS,YAAY,EAAE;KAC5C,MAAM,SAAS;KAChB,CAAC;;GAGP,CAAC;AAEF,OAAK,YAAY;AACjB,eAAa,EAAE,QAAQ,UAAU,SAAS,aAAa,CAAC,KAAK,CAAC;;;;;;;;;CAUhE,MAAM,WACJ,OACA,UAA6B,EAAE,EACJ;AAC3B,MAAI,CAAC,KAAK,UACR,OAAM,IAAI,MAAM,2CAA2C;EAG7D,MAAM,EAAE,UAAU,aAAa,OAAO,eAAe;EACrD,MAAM,YAAY,YAAY,KAAK;EAGnC,IAAIC;EACJ,IAAI,kBAAkB;AAEtB,MAAI,iBAAiB,YAAY;AAC/B,gBAAa,EAAE,QAAQ,qBAAqB,CAAC;GAC7C,MAAM,UAAU,UAAU,MAAM;AAChC,eAAY,QAAQ;AACpB,qBAAkB,QAAQ;QAE1B,aAAY;AAId,MAAI,oBAAoB,MAAO;AAC7B,gBAAa,EAAE,QAAQ,0BAA0B,CAAC;AAClD,eAAY,cAAc,WAAW,iBAAiB,KAAM;;EAG9D,MAAM,gBAAgB,UAAU,SAAS;AACzC,eAAa,EAAE,QAAQ,gBAAgB,cAAc,QAAQ,EAAE,CAAC,gBAAgB,CAAC;EAGjF,MAAMC,kBAAuB,EAAE;AAG/B,MAAI,YAAY,KAAK,YAAY,cAAc;AAC7C,mBAAgB,WAAW;AAC3B,mBAAgB,OAAO;;AAIzB,MAAI,WACF,iBAAgB,oBAAoB;EAItC,MAAM,SAAS,MAAM,KAAK,SAAS,WAAW,gBAAgB;EAE9D,MAAM,YAAY,YAAY,KAAK,GAAG;EAGtC,IAAI,OAAO,OAAO,MAAM,MAAM,IAAI;AAGlC,MAAI,SAAS,mBAAmB,SAAS,mBAAmB,SAAS,gBACnE,QAAO;EAGT,MAAMC,mBAAqC;GACzC;GACA,UAAU,aAAa,KAAK,YAAY,eAAe,SAAS;GAChE,UAAU;GACV;GACD;AAGD,MAAI,cAAc,OAAO,OACvB,kBAAiB,WAAW,OAAO,OAAO,KACvC,WAAmC;GAClC,MAAM,MAAM,MAAM,MAAM,IAAI;GAC5B,OAAO,MAAM,YAAY,MAAM;GAC/B,KAAK,MAAM,YAAY,MAAM;GAC9B,EACF;AAGH,eAAa,EAAE,QAAQ,SAAS,CAAC;AAEjC,SAAO;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8BT,uBACE,UAAyC,EAAE,EACZ;EAC/B,MAAM,EACJ,gBAAgB,KAChB,eAAe,KACf,SACA,cACA,SACA,aACE;EAEJ,IAAIC,cAA8B,EAAE;EACpC,IAAI,iBAAiB;EACrB,IAAI,aAAa;EACjB,IAAIC,aAAoD;EACxD,IAAI,YAAY;EAEhB,MAAM,sBAA8B;AAClC,UAAO,YAAY,QAAQ,KAAK,UAAU,MAAM,MAAM,QAAQ,EAAE;;EAGlE,MAAM,oBAAkC;GACtC,MAAM,cAAc,eAAe;GACnC,MAAM,SAAS,IAAI,aAAa,YAAY;GAC5C,IAAI,SAAS;AACb,QAAK,MAAM,SAAS,aAAa;AAC/B,WAAO,IAAI,OAAO,OAAO;AACzB,cAAU,MAAM;;AAElB,UAAO;;EAGT,MAAM,mBAAmB,YAA6B;AACpD,OAAI,CAAC,KAAK,aAAa,eAAe,GAAG,aACvC,QAAO;GAGT,MAAM,QAAQ,aAAa;AAC3B,iBAAc,EAAE;AAEhB,OAAI;IAEF,MAAM,QADS,MAAM,KAAK,WAAW,OAAO,EAAE,UAAU,CAAC,EACrC,KAAK,MAAM;AAE/B,QAAI,MAAM;AACR;AACA,eAAU,MAAM,WAAW;AAG3B,sBAAiB,kBAAkB,iBAAiB,MAAM,MAAM;AAChE,oBAAe,eAAe;;AAGhC,WAAO;YACAC,GAAQ;AACf,cAAU,EAAE,WAAW,uBAAuB;AAC9C,WAAO;;;EAIX,IAAI,UAAU;AAmEd,SAjE+C;GAC7C,YAAY,UAAwB;AAClC,QAAI,CAAC,QACH,aAAY,KAAK,MAAM;;GAI3B,OAAO,YAAY;AACjB,QAAI,QAAS,QAAO;AACpB,WAAO,kBAAkB;;GAG3B,aAAa;AACX,QAAI,aAAa,QAAS;AAC1B,gBAAY;AAEZ,iBAAa,YAAY,YAAY;AACnC,SAAI,aAAa,CAAC,QAChB,OAAM,kBAAkB;OAEzB,cAAc;;GAGnB,MAAM,YAAY;AAChB,gBAAY;AAEZ,QAAI,YAAY;AACd,mBAAc,WAAW;AACzB,kBAAa;;AAIf,QAAI,CAAC,WAAW,eAAe,IAAI,aACjC,OAAM,kBAAkB;AAG1B,WAAO;;GAGT,aAAa;AAEX,cAAU;AACV,gBAAY;AAEZ,QAAI,YAAY;AACd,mBAAc,WAAW;AACzB,kBAAa;;AAGf,kBAAc,EAAE;;GAGlB,iBAAiB;GAEjB,qBAAqB;GAErB,qBAAqB;GAErB,aAAa;AACX,kBAAc,EAAE;AAChB,qBAAiB;AACjB,iBAAa;;GAEhB;;;;;CAQH,UAAgB;AACd,OAAK,WAAW;AAChB,OAAK,YAAY;AACjB,OAAK,cAAc"}