@livekit/agents-plugin-silero 0.5.7 → 0.5.9

package/dist/vad.cjs

@@ -28,9 +28,10 @@ var import_onnx_model = require("./onnx_model.cjs");
 const SLOW_INFERENCE_THRESHOLD = 200;
 const defaultVADOptions = {
   minSpeechDuration: 50,
-  minSilenceDuration: 250,
+  minSilenceDuration: 550,
   prefixPaddingDuration: 500,
   maxBufferedSpeech: 6e4,
+  // 60 seconds
   activationThreshold: 0.5,
   sampleRate: 16e3,
   forceCPU: true
@@ -185,7 +186,7 @@ class VADStream extends import_agents.VADStream {
           const toCopyInt = Math.trunc(toCopy);
           inputCopyRemainingFrac = toCopy - toCopyInt;
           const availableSpace = this.#speechBuffer.length - speechBufferIndex;
-          const toCopyBuffer = Math.min(this.#model.windowSizeSamples, availableSpace);
+          const toCopyBuffer = Math.min(toCopyInt, availableSpace);
           if (toCopyBuffer > 0) {
             this.#speechBuffer.set(inputFrame.data.subarray(0, toCopyBuffer), speechBufferIndex);
             speechBufferIndex += toCopyBuffer;
@@ -204,9 +205,9 @@ class VADStream extends import_agents.VADStream {
             this.#logger.child({ delay: this.#extraInferenceTime }).warn("inference is slower than realtime");
           }
           if (pubSpeaking) {
-            pubSpeechDuration += inferenceDuration;
+            pubSpeechDuration += windowDuration;
           } else {
-            pubSilenceDuration += inferenceDuration;
+            pubSilenceDuration += windowDuration;
           }
           this.queue.put({
             type: import_agents.VADEventType.INFERENCE_DONE,

package/dist/vad.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/vad.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2024 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport {\n  ExpFilter,\n  VADEventType,\n  VADStream as baseStream,\n  VAD as baseVAD,\n  log,\n  mergeFrames,\n} from '@livekit/agents';\nimport { AudioFrame, AudioResampler, AudioResamplerQuality } from '@livekit/rtc-node';\nimport type { InferenceSession } from 'onnxruntime-node';\nimport type { SampleRate } from './onnx_model.js';\nimport { OnnxModel, newInferenceSession } from './onnx_model.js';\n\nconst SLOW_INFERENCE_THRESHOLD = 200; // late by 200ms\n\nexport interface VADOptions {\n  /** Minimum duration of speech to start a new speech chunk */\n  minSpeechDuration: number;\n  /** At the end of each speech, wait this duration before ending the speech */\n  minSilenceDuration: number;\n  /** Duration of padding to add to the beginning of each speech chunk */\n  prefixPaddingDuration: number;\n  /** Maximum duration of speech to keep in the buffer */\n  maxBufferedSpeech: number;\n  /** Maximum duration of speech to keep in the buffer*/\n  activationThreshold: number;\n  /** Sample rate for the inference (only 8KHz and 16KHz are supported) */\n  sampleRate: SampleRate;\n  /** Force the use of CPU for inference */\n  forceCPU: boolean;\n}\n\nconst defaultVADOptions: VADOptions = {\n  minSpeechDuration: 50,\n  minSilenceDuration: 250,\n  prefixPaddingDuration: 500,\n  maxBufferedSpeech: 60000,\n  activationThreshold: 0.5,\n  sampleRate: 16000,\n  forceCPU: true,\n};\n\nexport class VAD extends baseVAD {\n  #session: InferenceSession;\n  #opts: VADOptions;\n  #streams: VADStream[];\n  label = 'silero.VAD';\n\n  constructor(session: InferenceSession, opts: VADOptions) {\n    super({ updateInterval: 32 });\n    this.#session = session;\n    this.#opts = opts;\n    this.#streams = [];\n  }\n\n  /**\n   * Updates the VAD options with new values.\n   *\n   * @param opts - Partial options object containing the values to update\n   * @remarks\n   * This method will merge the provided options with existing options and update all active streams.\n   * Only the properties specified in opts will be updated, other properties retain their current values.\n   */\n  updateOptions(opts: Partial<VADOptions>): void {\n    this.#opts = { ...this.#opts, ...opts };\n    for (const stream of this.#streams) {\n      stream.updateOptions(this.#opts);\n    }\n  }\n\n  /**\n   * Load and initialize the Silero VAD model.\n   *\n   * This method loads the ONNX model and prepares it for inference. When options are not provided,\n   * sane defaults are used.\n   *\n   * @remarks\n   * This method may take time to load the model into memory.\n   * It is recommended to call this method inside your prewarm mechanism.\n   *\n   * @example\n   * ```ts\n   * export default defineAgent({\n   *   prewarm: async (proc: JobProcess) => {\n   *     proc.userData.vad = await VAD.load();\n   *   },\n   *   entry: async (ctx: JobContext) => {\n   *     const vad = ctx.proc.userData.vad! as VAD;\n   *     // the rest of your agent logic\n   *   },\n   * });\n   * ```\n   *\n   * @param options -\n   * @returns Promise\\<{@link VAD}\\>: An instance of the VAD class ready for streaming.\n   */\n  static async load(opts: Partial<VADOptions> = {}): Promise<VAD> {\n    const mergedOpts: VADOptions = { ...defaultVADOptions, ...opts };\n    const session = await newInferenceSession(mergedOpts.forceCPU);\n    return new VAD(session, mergedOpts);\n  }\n\n  stream(): VADStream {\n    const stream = new VADStream(\n      this,\n      this.#opts,\n      new OnnxModel(this.#session, this.#opts.sampleRate),\n    );\n    this.#streams.push(stream);\n    return stream;\n  }\n}\n\nexport class VADStream extends baseStream {\n  #opts: VADOptions;\n  #model: OnnxModel;\n  #inputSampleRate: number;\n  #speechBuffer: Int16Array | null;\n  #speechBufferMaxReached: boolean;\n  #prefixPaddingSamples: number;\n  #task: Promise<void>;\n  #expFilter = new ExpFilter(0.35);\n  #extraInferenceTime = 0;\n  #logger = log();\n\n  constructor(vad: VAD, opts: VADOptions, model: OnnxModel) {\n    super(vad);\n    this.#opts = opts;\n    this.#model = model;\n    this.#inputSampleRate = 0;\n    this.#speechBuffer = null;\n    this.#speechBufferMaxReached = false;\n    this.#prefixPaddingSamples = 0;\n\n    this.#task = new Promise(async () => {\n      let inferenceData = new Float32Array(this.#model.windowSizeSamples);\n\n      // a copy is exposed to the user in END_OF_SPEECH\n      let speechBufferIndex = 0;\n\n      // \"pub\" means public, these values are exposed to the users through events\n      let pubSpeaking = false;\n      let pubSpeechDuration = 0;\n      let pubSilenceDuration = 0;\n      let pubCurrentSample = 0;\n      let pubTimestamp = 0;\n      let speechThresholdDuration = 0;\n      let silenceThresholdDuration = 0;\n\n      let inputFrames = [];\n      let inferenceFrames: AudioFrame[] = [];\n      let resampler: AudioResampler | null = null;\n\n      // used to avoid drift when the sampleRate ratio is not an integer\n      let inputCopyRemainingFrac = 0.0;\n\n      for await (const frame of this.input) {\n        if (typeof frame === 'symbol') {\n          continue; // ignore flush sentinel for now\n        }\n\n        if (!this.#inputSampleRate || !this.#speechBuffer) {\n          this.#inputSampleRate = frame.sampleRate;\n          this.#prefixPaddingSamples = Math.trunc(\n            (this.#opts.prefixPaddingDuration * this.#inputSampleRate) / 1000,\n          );\n          const bufferSize =\n            Math.trunc((this.#opts.maxBufferedSpeech * this.#inputSampleRate) / 1000) +\n            this.#prefixPaddingSamples;\n          this.#speechBuffer = new Int16Array(bufferSize);\n\n          if (this.#opts.sampleRate !== this.#inputSampleRate) {\n            // resampling needed: the input sample rate isn't the same as the model's\n            // sample rate used for inference\n            resampler = new AudioResampler(\n              this.#inputSampleRate,\n              this.#opts.sampleRate,\n              1,\n              AudioResamplerQuality.QUICK, // VAD doesn't need high quality\n            );\n          }\n        } else if (frame.sampleRate !== this.#inputSampleRate) {\n          this.#logger.error('a frame with a different sample rate was already published');\n          continue;\n        }\n\n        inputFrames.push(frame);\n        if (resampler) {\n          inferenceFrames.push(...resampler.push(frame));\n        } else {\n          inferenceFrames.push(frame);\n        }\n\n        while (true) {\n          const startTime = process.hrtime.bigint();\n          const availableInferenceSamples = inferenceFrames\n            .map((x) => x.samplesPerChannel)\n            .reduce((acc, x) => acc + x, 0);\n\n          if (availableInferenceSamples < this.#model.windowSizeSamples) {\n            break; // not enough samples to run inference\n          }\n\n          const inputFrame = mergeFrames(inputFrames);\n          const inferenceFrame = mergeFrames(inferenceFrames);\n\n          // convert data to f32\n          inferenceData = Float32Array.from(\n            inferenceFrame.data.subarray(0, this.#model.windowSizeSamples),\n            (x) => x / 32767,\n          );\n\n          const p = await this.#model\n            .run(inferenceData)\n            .then((data) => this.#expFilter.apply(1, data));\n\n          const windowDuration = (this.#model.windowSizeSamples / this.#opts.sampleRate) * 1000;\n          pubCurrentSample += this.#model.windowSizeSamples;\n          pubTimestamp += windowDuration;\n          const resamplingRatio = this.#inputSampleRate / this.#model.sampleRate;\n          const toCopy = this.#model.windowSizeSamples * resamplingRatio + inputCopyRemainingFrac;\n          const toCopyInt = Math.trunc(toCopy);\n          inputCopyRemainingFrac = toCopy - toCopyInt;\n\n          // copy the inference window to the speech buffer\n          const availableSpace = this.#speechBuffer.length - speechBufferIndex;\n          const toCopyBuffer = Math.min(this.#model.windowSizeSamples, availableSpace);\n          if (toCopyBuffer > 0) {\n            this.#speechBuffer.set(inputFrame.data.subarray(0, toCopyBuffer), speechBufferIndex);\n            speechBufferIndex += toCopyBuffer;\n          } else if (!this.#speechBufferMaxReached) {\n            this.#speechBufferMaxReached = true;\n            this.#logger.warn(\n              'maxBufferedSpeech reached, ignoring further data for the current speech input',\n            );\n          }\n\n          const inferenceDuration = Number((process.hrtime.bigint() - startTime) / BigInt(1000000));\n          this.#extraInferenceTime = Math.max(\n            0,\n            this.#extraInferenceTime + inferenceDuration - windowDuration,\n          );\n          if (this.#extraInferenceTime > SLOW_INFERENCE_THRESHOLD) {\n            this.#logger\n              .child({ delay: this.#extraInferenceTime })\n              .warn('inference is slower than realtime');\n          }\n\n          if (pubSpeaking) {\n            pubSpeechDuration += inferenceDuration;\n          } else {\n            pubSilenceDuration += inferenceDuration;\n          }\n\n          this.queue.put({\n            type: VADEventType.INFERENCE_DONE,\n            samplesIndex: pubCurrentSample,\n            timestamp: pubTimestamp,\n            silenceDuration: pubSilenceDuration,\n            speechDuration: pubSpeechDuration,\n            probability: p,\n            inferenceDuration,\n            frames: [\n              new AudioFrame(\n                inputFrame.data.subarray(0, toCopyInt),\n                this.#inputSampleRate,\n                1,\n                toCopyInt,\n              ),\n            ],\n            speaking: pubSpeaking,\n            rawAccumulatedSilence: silenceThresholdDuration,\n            rawAccumulatedSpeech: speechThresholdDuration,\n          });\n\n          const resetWriteCursor = () => {\n            if (!this.#speechBuffer) throw new Error('speechBuffer is empty');\n            if (speechBufferIndex <= this.#prefixPaddingSamples) {\n              return;\n            }\n\n            const paddingData = this.#speechBuffer.subarray(\n              speechBufferIndex - this.#prefixPaddingSamples,\n              speechBufferIndex,\n            );\n            this.#speechBuffer.set(paddingData, 0);\n            speechBufferIndex = this.#prefixPaddingSamples;\n            this.#speechBufferMaxReached = false;\n          };\n\n          const copySpeechBuffer = (): AudioFrame => {\n            if (!this.#speechBuffer) throw new Error('speechBuffer is empty');\n            return new AudioFrame(\n              this.#speechBuffer.subarray(this.#prefixPaddingSamples, speechBufferIndex),\n              this.#inputSampleRate,\n              1,\n              speechBufferIndex,\n            );\n          };\n\n          if (p > this.#opts.activationThreshold) {\n            speechThresholdDuration += windowDuration;\n            silenceThresholdDuration = 0;\n            if (!pubSpeaking && speechThresholdDuration >= this.#opts.minSpeechDuration) {\n              pubSpeaking = true;\n              pubSilenceDuration = 0;\n              pubSpeechDuration = speechThresholdDuration;\n\n              this.queue.put({\n                type: VADEventType.START_OF_SPEECH,\n                samplesIndex: pubCurrentSample,\n                timestamp: pubTimestamp,\n                silenceDuration: pubSilenceDuration,\n                speechDuration: pubSpeechDuration,\n                probability: p,\n                inferenceDuration,\n                frames: [copySpeechBuffer()],\n                speaking: pubSpeaking,\n                rawAccumulatedSilence: 0,\n                rawAccumulatedSpeech: 0,\n              });\n            }\n          } else {\n            silenceThresholdDuration += windowDuration;\n            speechThresholdDuration = 0;\n\n            if (!pubSpeaking) {\n              resetWriteCursor();\n            }\n\n            if (pubSpeaking && silenceThresholdDuration > this.#opts.minSilenceDuration) {\n              pubSpeaking = false;\n              pubSpeechDuration = 0;\n              pubSilenceDuration = silenceThresholdDuration;\n\n              this.queue.put({\n                type: VADEventType.END_OF_SPEECH,\n                samplesIndex: pubCurrentSample,\n                timestamp: pubTimestamp,\n                silenceDuration: pubSilenceDuration,\n                speechDuration: pubSpeechDuration,\n                probability: p,\n                inferenceDuration,\n                frames: [copySpeechBuffer()],\n                speaking: pubSpeaking,\n                rawAccumulatedSilence: 0,\n                rawAccumulatedSpeech: 0,\n              });\n\n              resetWriteCursor();\n            }\n          }\n\n          inputFrames = [];\n          inferenceFrames = [];\n\n          if (inputFrame.data.length > toCopyInt) {\n            const data = inputFrame.data.subarray(toCopyInt);\n            inputFrames.push(\n              new AudioFrame(data, this.#inputSampleRate, 1, Math.trunc(data.length / 2)),\n            );\n          }\n          if (inferenceFrame.data.length > this.#model.windowSizeSamples) {\n            const data = inferenceFrame.data.subarray(this.#model.windowSizeSamples);\n            inferenceFrames.push(\n              new AudioFrame(data, this.#opts.sampleRate, 1, Math.trunc(data.length / 2)),\n            );\n          }\n        }\n      }\n    });\n  }\n\n  /**\n   * Update the VAD options\n   *\n   * @param opts - Partial options object containing the values to update\n   * @remarks\n   * This method allows you to update the VAD options after the VAD object has been created\n   */\n  updateOptions(opts: Partial<VADOptions>) {\n    const oldMaxBufferedSpeech = this.#opts.maxBufferedSpeech;\n    this.#opts = { ...this.#opts, ...opts };\n\n    if (this.#inputSampleRate) {\n      // Assert speech buffer exists\n      if (this.#speechBuffer === null) throw new Error('speechBuffer is null');\n\n      // Resize speech buffer\n      this.#prefixPaddingSamples = Math.trunc(\n        (this.#opts.prefixPaddingDuration * this.#inputSampleRate) / 1000,\n      );\n      const bufferSize =\n        Math.trunc((this.#opts.maxBufferedSpeech * this.#inputSampleRate) / 1000) +\n        this.#prefixPaddingSamples;\n      const resizedBuffer = new Int16Array(bufferSize);\n      resizedBuffer.set(\n        this.#speechBuffer.subarray(0, Math.min(this.#speechBuffer.length, bufferSize)),\n      );\n      this.#speechBuffer = resizedBuffer;\n\n      // Determine if max has been reached\n      if (this.#opts.maxBufferedSpeech > oldMaxBufferedSpeech) {\n        this.#speechBufferMaxReached = false;\n      }\n    }\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAGA,oBAOO;AACP,sBAAkE;AAGlE,wBAA+C;AAE/C,MAAM,2BAA2B;AAmBjC,MAAM,oBAAgC;AAAA,EACpC,mBAAmB;AAAA,EACnB,oBAAoB;AAAA,EACpB,uBAAuB;AAAA,EACvB,mBAAmB;AAAA,EACnB,qBAAqB;AAAA,EACrB,YAAY;AAAA,EACZ,UAAU;AACZ;AAEO,MAAM,YAAY,cAAAA,IAAQ;AAAA,EAC/B;AAAA,EACA;AAAA,EACA;AAAA,EACA,QAAQ;AAAA,EAER,YAAY,SAA2B,MAAkB;AACvD,UAAM,EAAE,gBAAgB,GAAG,CAAC;AAC5B,SAAK,WAAW;AAChB,SAAK,QAAQ;AACb,SAAK,WAAW,CAAC;AAAA,EACnB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUA,cAAc,MAAiC;AAC7C,SAAK,QAAQ,EAAE,GAAG,KAAK,OAAO,GAAG,KAAK;AACtC,eAAW,UAAU,KAAK,UAAU;AAClC,aAAO,cAAc,KAAK,KAAK;AAAA,IACjC;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EA4BA,aAAa,KAAK,OAA4B,CAAC,GAAiB;AAC9D,UAAM,aAAyB,EAAE,GAAG,mBAAmB,GAAG,KAAK;AAC/D,UAAM,UAAU,UAAM,uCAAoB,WAAW,QAAQ;AAC7D,WAAO,IAAI,IAAI,SAAS,UAAU;AAAA,EACpC;AAAA,EAEA,SAAoB;AAClB,UAAM,SAAS,IAAI;AAAA,MACjB;AAAA,MACA,KAAK;AAAA,MACL,IAAI,4BAAU,KAAK,UAAU,KAAK,MAAM,UAAU;AAAA,IACpD;AACA,SAAK,SAAS,KAAK,MAAM;AACzB,WAAO;AAAA,EACT;AACF;AAEO,MAAM,kBAAkB,cAAAC,UAAW;AAAA,EACxC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA,aAAa,IAAI,wBAAU,IAAI;AAAA,EAC/B,sBAAsB;AAAA,EACtB,cAAU,mBAAI;AAAA,EAEd,YAAY,KAAU,MAAkB,OAAkB;AACxD,UAAM,GAAG;AACT,SAAK,QAAQ;AACb,SAAK,SAAS;AACd,SAAK,mBAAmB;AACxB,SAAK,gBAAgB;AACrB,SAAK,0BAA0B;AAC/B,SAAK,wBAAwB;AAE7B,SAAK,QAAQ,IAAI,QAAQ,YAAY;AACnC,UAAI,gBAAgB,IAAI,aAAa,KAAK,OAAO,iBAAiB;AAGlE,UAAI,oBAAoB;AAGxB,UAAI,cAAc;AAClB,UAAI,oBAAoB;AACxB,UAAI,qBAAqB;AACzB,UAAI,mBAAmB;AACvB,UAAI,eAAe;AACnB,UAAI,0BAA0B;AAC9B,UAAI,2BAA2B;AAE/B,UAAI,cAAc,CAAC;AACnB,UAAI,kBAAgC,CAAC;AACrC,UAAI,YAAmC;AAGvC,UAAI,yBAAyB;AAE7B,uBAAiB,SAAS,KAAK,OAAO;AACpC,YAAI,OAAO,UAAU,UAAU;AAC7B;AAAA,QACF;AAEA,YAAI,CAAC,KAAK,oBAAoB,CAAC,KAAK,eAAe;AACjD,eAAK,mBAAmB,MAAM;AAC9B,eAAK,wBAAwB,KAAK;AAAA,YAC/B,KAAK,MAAM,wBAAwB,KAAK,mBAAoB;AAAA,UAC/D;AACA,gBAAM,aACJ,KAAK,MAAO,KAAK,MAAM,oBAAoB,KAAK,mBAAoB,GAAI,IACxE,KAAK;AACP,eAAK,gBAAgB,IAAI,WAAW,UAAU;AAE9C,cAAI,KAAK,MAAM,eAAe,KAAK,kBAAkB;AAGnD,wBAAY,IAAI;AAAA,cACd,KAAK;AAAA,cACL,KAAK,MAAM;AAAA,cACX;AAAA,cACA,sCAAsB;AAAA;AAAA,YACxB;AAAA,UACF;AAAA,QACF,WAAW,MAAM,eAAe,KAAK,kBAAkB;AACrD,eAAK,QAAQ,MAAM,4DAA4D;AAC/E;AAAA,QACF;AAEA,oBAAY,KAAK,KAAK;AACtB,YAAI,WAAW;AACb,0BAAgB,KAAK,GAAG,UAAU,KAAK,KAAK,CAAC;AAAA,QAC/C,OAAO;AACL,0BAAgB,KAAK,KAAK;AAAA,QAC5B;AAEA,eAAO,MAAM;AACX,gBAAM,YAAY,QAAQ,OAAO,OAAO;AACxC,gBAAM,4BAA4B,gBAC/B,IAAI,CAAC,MAAM,EAAE,iBAAiB,EAC9B,OAAO,CAAC,KAAK,MAAM,MAAM,GAAG,CAAC;AAEhC,cAAI,4BAA4B,KAAK,OAAO,mBAAmB;AAC7D;AAAA,UACF;AAEA,gBAAM,iBAAa,2BAAY,WAAW;AAC1C,gBAAM,qBAAiB,2BAAY,eAAe;AAGlD,0BAAgB,aAAa;AAAA,YAC3B,eAAe,KAAK,SAAS,GAAG,KAAK,OAAO,iBAAiB;AAAA,YAC7D,CAAC,MAAM,IAAI;AAAA,UACb;AAEA,gBAAM,IAAI,MAAM,KAAK,OAClB,IAAI,aAAa,EACjB,KAAK,CAAC,SAAS,KAAK,WAAW,MAAM,GAAG,IAAI,CAAC;AAEhD,gBAAM,iBAAkB,KAAK,OAAO,oBAAoB,KAAK,MAAM,aAAc;AACjF,8BAAoB,KAAK,OAAO;AAChC,0BAAgB;AAChB,gBAAM,kBAAkB,KAAK,mBAAmB,KAAK,OAAO;AAC5D,gBAAM,SAAS,KAAK,OAAO,oBAAoB,kBAAkB;AACjE,gBAAM,YAAY,KAAK,MAAM,MAAM;AACnC,mCAAyB,SAAS;AAGlC,gBAAM,iBAAiB,KAAK,cAAc,SAAS;AACnD,gBAAM,eAAe,KAAK,IAAI,KAAK,OAAO,mBAAmB,cAAc;AAC3E,cAAI,eAAe,GAAG;AACpB,iBAAK,cAAc,IAAI,WAAW,KAAK,SAAS,GAAG,YAAY,GAAG,iBAAiB;AACnF,iCAAqB;AAAA,UACvB,WAAW,CAAC,KAAK,yBAAyB;AACxC,iBAAK,0BAA0B;AAC/B,iBAAK,QAAQ;AAAA,cACX;AAAA,YACF;AAAA,UACF;AAEA,gBAAM,oBAAoB,QAAQ,QAAQ,OAAO,OAAO,IAAI,aAAa,OAAO,GAAO,CAAC;AACxF,eAAK,sBAAsB,KAAK;AAAA,YAC9B;AAAA,YACA,KAAK,sBAAsB,oBAAoB;AAAA,UACjD;AACA,cAAI,KAAK,sBAAsB,0BAA0B;AACvD,iBAAK,QACF,MAAM,EAAE,OAAO,KAAK,oBAAoB,CAAC,EACzC,KAAK,mCAAmC;AAAA,UAC7C;AAEA,cAAI,aAAa;AACf,iCAAqB;AAAA,UACvB,OAAO;AACL,kCAAsB;AAAA,UACxB;AAEA,eAAK,MAAM,IAAI;AAAA,YACb,MAAM,2BAAa;AAAA,YACnB,cAAc;AAAA,YACd,WAAW;AAAA,YACX,iBAAiB;AAAA,YACjB,gBAAgB;AAAA,YAChB,aAAa;AAAA,YACb;AAAA,YACA,QAAQ;AAAA,cACN,IAAI;AAAA,gBACF,WAAW,KAAK,SAAS,GAAG,SAAS;AAAA,gBACrC,KAAK;AAAA,gBACL;AAAA,gBACA;AAAA,cACF;AAAA,YACF;AAAA,YACA,UAAU;AAAA,YACV,uBAAuB;AAAA,YACvB,sBAAsB;AAAA,UACxB,CAAC;AAED,gBAAM,mBAAmB,MAAM;AAC7B,gBAAI,CAAC,KAAK,cAAe,OAAM,IAAI,MAAM,uBAAuB;AAChE,gBAAI,qBAAqB,KAAK,uBAAuB;AACnD;AAAA,YACF;AAEA,kBAAM,cAAc,KAAK,cAAc;AAAA,cACrC,oBAAoB,KAAK;AAAA,cACzB;AAAA,YACF;AACA,iBAAK,cAAc,IAAI,aAAa,CAAC;AACrC,gCAAoB,KAAK;AACzB,iBAAK,0BAA0B;AAAA,UACjC;AAEA,gBAAM,mBAAmB,MAAkB;AACzC,gBAAI,CAAC,KAAK,cAAe,OAAM,IAAI,MAAM,uBAAuB;AAChE,mBAAO,IAAI;AAAA,cACT,KAAK,cAAc,SAAS,KAAK,uBAAuB,iBAAiB;AAAA,cACzE,KAAK;AAAA,cACL;AAAA,cACA;AAAA,YACF;AAAA,UACF;AAEA,cAAI,IAAI,KAAK,MAAM,qBAAqB;AACtC,uCAA2B;AAC3B,uCAA2B;AAC3B,gBAAI,CAAC,eAAe,2BAA2B,KAAK,MAAM,mBAAmB;AAC3E,4BAAc;AACd,mCAAqB;AACrB,kCAAoB;AAEpB,mBAAK,MAAM,IAAI;AAAA,gBACb,MAAM,2BAAa;AAAA,gBACnB,cAAc;AAAA,gBACd,WAAW;AAAA,gBACX,iBAAiB;AAAA,gBACjB,gBAAgB;AAAA,gBAChB,aAAa;AAAA,gBACb;AAAA,gBACA,QAAQ,CAAC,iBAAiB,CAAC;AAAA,gBAC3B,UAAU;AAAA,gBACV,uBAAuB;AAAA,gBACvB,sBAAsB;AAAA,cACxB,CAAC;AAAA,YACH;AAAA,UACF,OAAO;AACL,wCAA4B;AAC5B,sCAA0B;AAE1B,gBAAI,CAAC,aAAa;AAChB,+BAAiB;AAAA,YACnB;AAEA,gBAAI,eAAe,2BAA2B,KAAK,MAAM,oBAAoB;AAC3E,4BAAc;AACd,kCAAoB;AACpB,mCAAqB;AAErB,mBAAK,MAAM,IAAI;AAAA,gBACb,MAAM,2BAAa;AAAA,gBACnB,cAAc;AAAA,gBACd,WAAW;AAAA,gBACX,iBAAiB;AAAA,gBACjB,gBAAgB;AAAA,gBAChB,aAAa;AAAA,gBACb;AAAA,gBACA,QAAQ,CAAC,iBAAiB,CAAC;AAAA,gBAC3B,UAAU;AAAA,gBACV,uBAAuB;AAAA,gBACvB,sBAAsB;AAAA,cACxB,CAAC;AAED,+BAAiB;AAAA,YACnB;AAAA,UACF;AAEA,wBAAc,CAAC;AACf,4BAAkB,CAAC;AAEnB,cAAI,WAAW,KAAK,SAAS,WAAW;AACtC,kBAAM,OAAO,WAAW,KAAK,SAAS,SAAS;AAC/C,wBAAY;AAAA,cACV,IAAI,2BAAW,MAAM,KAAK,kBAAkB,GAAG,KAAK,MAAM,KAAK,SAAS,CAAC,CAAC;AAAA,YAC5E;AAAA,UACF;AACA,cAAI,eAAe,KAAK,SAAS,KAAK,OAAO,mBAAmB;AAC9D,kBAAM,OAAO,eAAe,KAAK,SAAS,KAAK,OAAO,iBAAiB;AACvE,4BAAgB;AAAA,cACd,IAAI,2BAAW,MAAM,KAAK,MAAM,YAAY,GAAG,KAAK,MAAM,KAAK,SAAS,CAAC,CAAC;AAAA,YAC5E;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAAA,IACF,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,cAAc,MAA2B;AACvC,UAAM,uBAAuB,KAAK,MAAM;AACxC,SAAK,QAAQ,EAAE,GAAG,KAAK,OAAO,GAAG,KAAK;AAEtC,QAAI,KAAK,kBAAkB;AAEzB,UAAI,KAAK,kBAAkB,KAAM,OAAM,IAAI,MAAM,sBAAsB;AAGvE,WAAK,wBAAwB,KAAK;AAAA,QAC/B,KAAK,MAAM,wBAAwB,KAAK,mBAAoB;AAAA,MAC/D;AACA,YAAM,aACJ,KAAK,MAAO,KAAK,MAAM,oBAAoB,KAAK,mBAAoB,GAAI,IACxE,KAAK;AACP,YAAM,gBAAgB,IAAI,WAAW,UAAU;AAC/C,oBAAc;AAAA,QACZ,KAAK,cAAc,SAAS,GAAG,KAAK,IAAI,KAAK,cAAc,QAAQ,UAAU,CAAC;AAAA,MAChF;AACA,WAAK,gBAAgB;AAGrB,UAAI,KAAK,MAAM,oBAAoB,sBAAsB;AACvD,aAAK,0BAA0B;AAAA,MACjC;AAAA,IACF;AAAA,EACF;AACF;","names":["baseVAD","baseStream"]}
+{"version":3,"sources":["../src/vad.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2024 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport {\n  ExpFilter,\n  VADEventType,\n  VADStream as baseStream,\n  VAD as baseVAD,\n  log,\n  mergeFrames,\n} from '@livekit/agents';\nimport { AudioFrame, AudioResampler, AudioResamplerQuality } from '@livekit/rtc-node';\nimport type { InferenceSession } from 'onnxruntime-node';\nimport type { SampleRate } from './onnx_model.js';\nimport { OnnxModel, newInferenceSession } from './onnx_model.js';\n\nconst SLOW_INFERENCE_THRESHOLD = 200; // late by 200ms\n\nexport interface VADOptions {\n  /** Minimum duration of speech to start a new speech chunk */\n  minSpeechDuration: number;\n  /** At the end of each speech, wait this duration before ending the speech */\n  minSilenceDuration: number;\n  /** Duration of padding to add to the beginning of each speech chunk */\n  prefixPaddingDuration: number;\n  /** Maximum duration of speech to keep in the buffer */\n  maxBufferedSpeech: number;\n  /** Maximum duration of speech to keep in the buffer*/\n  activationThreshold: number;\n  /** Sample rate for the inference (only 8KHz and 16KHz are supported) */\n  sampleRate: SampleRate;\n  /** Force the use of CPU for inference */\n  forceCPU: boolean;\n}\n\nconst defaultVADOptions: VADOptions = {\n  minSpeechDuration: 50,\n  minSilenceDuration: 550,\n  prefixPaddingDuration: 500,\n  maxBufferedSpeech: 60000, // 60 seconds\n  activationThreshold: 0.5,\n  sampleRate: 16000,\n  forceCPU: true,\n};\n\nexport class VAD extends baseVAD {\n  #session: InferenceSession;\n  #opts: VADOptions;\n  #streams: VADStream[];\n  label = 'silero.VAD';\n\n  constructor(session: InferenceSession, opts: VADOptions) {\n    super({ updateInterval: 32 });\n    this.#session = session;\n    this.#opts = opts;\n    this.#streams = [];\n  }\n\n  /**\n   * Updates the VAD options with new values.\n   *\n   * @param opts - Partial options object containing the values to update\n   * @remarks\n   * This method will merge the provided options with existing options and update all active streams.\n   * Only the properties specified in opts will be updated, other properties retain their current values.\n   */\n  updateOptions(opts: Partial<VADOptions>): void {\n    this.#opts = { ...this.#opts, ...opts };\n    for (const stream of this.#streams) {\n      stream.updateOptions(this.#opts);\n    }\n  }\n\n  /**\n   * Load and initialize the Silero VAD model.\n   *\n   * This method loads the ONNX model and prepares it for inference. When options are not provided,\n   * sane defaults are used.\n   *\n   * @remarks\n   * This method may take time to load the model into memory.\n   * It is recommended to call this method inside your prewarm mechanism.\n   *\n   * @example\n   * ```ts\n   * export default defineAgent({\n   *   prewarm: async (proc: JobProcess) => {\n   *     proc.userData.vad = await VAD.load();\n   *   },\n   *   entry: async (ctx: JobContext) => {\n   *     const vad = ctx.proc.userData.vad! as VAD;\n   *     // the rest of your agent logic\n   *   },\n   * });\n   * ```\n   *\n   * @param options -\n   * @returns Promise\\<{@link VAD}\\>: An instance of the VAD class ready for streaming.\n   */\n  static async load(opts: Partial<VADOptions> = {}): Promise<VAD> {\n    const mergedOpts: VADOptions = { ...defaultVADOptions, ...opts };\n    const session = await newInferenceSession(mergedOpts.forceCPU);\n    return new VAD(session, mergedOpts);\n  }\n\n  stream(): VADStream {\n    const stream = new VADStream(\n      this,\n      this.#opts,\n      new OnnxModel(this.#session, this.#opts.sampleRate),\n    );\n    this.#streams.push(stream);\n    return stream;\n  }\n}\n\nexport class VADStream extends baseStream {\n  #opts: VADOptions;\n  #model: OnnxModel;\n  #inputSampleRate: number;\n  #speechBuffer: Int16Array | null;\n  #speechBufferMaxReached: boolean;\n  #prefixPaddingSamples: number;\n  #task: Promise<void>;\n  #expFilter = new ExpFilter(0.35);\n  #extraInferenceTime = 0;\n  #logger = log();\n\n  constructor(vad: VAD, opts: VADOptions, model: OnnxModel) {\n    super(vad);\n    this.#opts = opts;\n    this.#model = model;\n    this.#inputSampleRate = 0;\n    this.#speechBuffer = null;\n    this.#speechBufferMaxReached = false;\n    this.#prefixPaddingSamples = 0;\n\n    this.#task = new Promise(async () => {\n      let inferenceData = new Float32Array(this.#model.windowSizeSamples);\n\n      // a copy is exposed to the user in END_OF_SPEECH\n      let speechBufferIndex = 0;\n\n      // \"pub\" means public, these values are exposed to the users through events\n      let pubSpeaking = false;\n      let pubSpeechDuration = 0;\n      let pubSilenceDuration = 0;\n      let pubCurrentSample = 0;\n      let pubTimestamp = 0;\n      let speechThresholdDuration = 0;\n      let silenceThresholdDuration = 0;\n\n      let inputFrames = [];\n      let inferenceFrames: AudioFrame[] = [];\n      let resampler: AudioResampler | null = null;\n\n      // used to avoid drift when the sampleRate ratio is not an integer\n      let inputCopyRemainingFrac = 0.0;\n\n      for await (const frame of this.input) {\n        if (typeof frame === 'symbol') {\n          continue; // ignore flush sentinel for now\n        }\n\n        if (!this.#inputSampleRate || !this.#speechBuffer) {\n          this.#inputSampleRate = frame.sampleRate;\n          this.#prefixPaddingSamples = Math.trunc(\n            (this.#opts.prefixPaddingDuration * this.#inputSampleRate) / 1000,\n          );\n          const bufferSize =\n            Math.trunc((this.#opts.maxBufferedSpeech * this.#inputSampleRate) / 1000) +\n            this.#prefixPaddingSamples;\n          this.#speechBuffer = new Int16Array(bufferSize);\n\n          if (this.#opts.sampleRate !== this.#inputSampleRate) {\n            // resampling needed: the input sample rate isn't the same as the model's\n            // sample rate used for inference\n            resampler = new AudioResampler(\n              this.#inputSampleRate,\n              this.#opts.sampleRate,\n              1,\n              AudioResamplerQuality.QUICK, // VAD doesn't need high quality\n            );\n          }\n        } else if (frame.sampleRate !== this.#inputSampleRate) {\n          this.#logger.error('a frame with a different sample rate was already published');\n          continue;\n        }\n\n        inputFrames.push(frame);\n        if (resampler) {\n          inferenceFrames.push(...resampler.push(frame));\n        } else {\n          inferenceFrames.push(frame);\n        }\n\n        while (true) {\n          const startTime = process.hrtime.bigint();\n          const availableInferenceSamples = inferenceFrames\n            .map((x) => x.samplesPerChannel)\n            .reduce((acc, x) => acc + x, 0);\n\n          if (availableInferenceSamples < this.#model.windowSizeSamples) {\n            break; // not enough samples to run inference\n          }\n\n          const inputFrame = mergeFrames(inputFrames);\n          const inferenceFrame = mergeFrames(inferenceFrames);\n\n          // convert data to f32\n          inferenceData = Float32Array.from(\n            inferenceFrame.data.subarray(0, this.#model.windowSizeSamples),\n            (x) => x / 32767,\n          );\n\n          const p = await this.#model\n            .run(inferenceData)\n            .then((data) => this.#expFilter.apply(1, data));\n\n          const windowDuration = (this.#model.windowSizeSamples / this.#opts.sampleRate) * 1000;\n          pubCurrentSample += this.#model.windowSizeSamples;\n          pubTimestamp += windowDuration;\n          const resamplingRatio = this.#inputSampleRate / this.#model.sampleRate;\n          const toCopy = this.#model.windowSizeSamples * resamplingRatio + inputCopyRemainingFrac;\n          const toCopyInt = Math.trunc(toCopy);\n          inputCopyRemainingFrac = toCopy - toCopyInt;\n\n          // copy the inference window to the speech buffer\n          const availableSpace = this.#speechBuffer.length - speechBufferIndex;\n          const toCopyBuffer = Math.min(toCopyInt, availableSpace);\n          if (toCopyBuffer > 0) {\n            this.#speechBuffer.set(inputFrame.data.subarray(0, toCopyBuffer), speechBufferIndex);\n            speechBufferIndex += toCopyBuffer;\n          } else if (!this.#speechBufferMaxReached) {\n            this.#speechBufferMaxReached = true;\n            this.#logger.warn(\n              'maxBufferedSpeech reached, ignoring further data for the current speech input',\n            );\n          }\n\n          const inferenceDuration = Number((process.hrtime.bigint() - startTime) / BigInt(1000000));\n          this.#extraInferenceTime = Math.max(\n            0,\n            this.#extraInferenceTime + inferenceDuration - windowDuration,\n          );\n          if (this.#extraInferenceTime > SLOW_INFERENCE_THRESHOLD) {\n            this.#logger\n              .child({ delay: this.#extraInferenceTime })\n              .warn('inference is slower than realtime');\n          }\n\n          if (pubSpeaking) {\n            pubSpeechDuration += windowDuration;\n          } else {\n            pubSilenceDuration += windowDuration;\n          }\n\n          this.queue.put({\n            type: VADEventType.INFERENCE_DONE,\n            samplesIndex: pubCurrentSample,\n            timestamp: pubTimestamp,\n            silenceDuration: pubSilenceDuration,\n            speechDuration: pubSpeechDuration,\n            probability: p,\n            inferenceDuration,\n            frames: [\n              new AudioFrame(\n                inputFrame.data.subarray(0, toCopyInt),\n                this.#inputSampleRate,\n                1,\n                toCopyInt,\n              ),\n            ],\n            speaking: pubSpeaking,\n            rawAccumulatedSilence: silenceThresholdDuration,\n            rawAccumulatedSpeech: speechThresholdDuration,\n          });\n\n          const resetWriteCursor = () => {\n            if (!this.#speechBuffer) throw new Error('speechBuffer is empty');\n            if (speechBufferIndex <= this.#prefixPaddingSamples) {\n              return;\n            }\n\n            const paddingData = this.#speechBuffer.subarray(\n              speechBufferIndex - this.#prefixPaddingSamples,\n              speechBufferIndex,\n            );\n            this.#speechBuffer.set(paddingData, 0);\n            speechBufferIndex = this.#prefixPaddingSamples;\n            this.#speechBufferMaxReached = false;\n          };\n\n          const copySpeechBuffer = (): AudioFrame => {\n            if (!this.#speechBuffer) throw new Error('speechBuffer is empty');\n            return new AudioFrame(\n              this.#speechBuffer.subarray(this.#prefixPaddingSamples, speechBufferIndex),\n              this.#inputSampleRate,\n              1,\n              speechBufferIndex,\n            );\n          };\n\n          if (p > this.#opts.activationThreshold) {\n            speechThresholdDuration += windowDuration;\n            silenceThresholdDuration = 0;\n            if (!pubSpeaking && speechThresholdDuration >= this.#opts.minSpeechDuration) {\n              pubSpeaking = true;\n              pubSilenceDuration = 0;\n              pubSpeechDuration = speechThresholdDuration;\n\n              this.queue.put({\n                type: VADEventType.START_OF_SPEECH,\n                samplesIndex: pubCurrentSample,\n                timestamp: pubTimestamp,\n                silenceDuration: pubSilenceDuration,\n                speechDuration: pubSpeechDuration,\n                probability: p,\n                inferenceDuration,\n                frames: [copySpeechBuffer()],\n                speaking: pubSpeaking,\n                rawAccumulatedSilence: 0,\n                rawAccumulatedSpeech: 0,\n              });\n            }\n          } else {\n            silenceThresholdDuration += windowDuration;\n            speechThresholdDuration = 0;\n\n            if (!pubSpeaking) {\n              resetWriteCursor();\n            }\n\n            if (pubSpeaking && silenceThresholdDuration > this.#opts.minSilenceDuration) {\n              pubSpeaking = false;\n              pubSpeechDuration = 0;\n              pubSilenceDuration = silenceThresholdDuration;\n\n              this.queue.put({\n                type: VADEventType.END_OF_SPEECH,\n                samplesIndex: pubCurrentSample,\n                timestamp: pubTimestamp,\n                silenceDuration: pubSilenceDuration,\n                speechDuration: pubSpeechDuration,\n                probability: p,\n                inferenceDuration,\n                frames: [copySpeechBuffer()],\n                speaking: pubSpeaking,\n                rawAccumulatedSilence: 0,\n                rawAccumulatedSpeech: 0,\n              });\n\n              resetWriteCursor();\n            }\n          }\n\n          inputFrames = [];\n          inferenceFrames = [];\n\n          if (inputFrame.data.length > toCopyInt) {\n            const data = inputFrame.data.subarray(toCopyInt);\n            inputFrames.push(\n              new AudioFrame(data, this.#inputSampleRate, 1, Math.trunc(data.length / 2)),\n            );\n          }\n          if (inferenceFrame.data.length > this.#model.windowSizeSamples) {\n            const data = inferenceFrame.data.subarray(this.#model.windowSizeSamples);\n            inferenceFrames.push(\n              new AudioFrame(data, this.#opts.sampleRate, 1, Math.trunc(data.length / 2)),\n            );\n          }\n        }\n      }\n    });\n  }\n\n  /**\n   * Update the VAD options\n   *\n   * @param opts - Partial options object containing the values to update\n   * @remarks\n   * This method allows you to update the VAD options after the VAD object has been created\n   */\n  updateOptions(opts: Partial<VADOptions>) {\n    const oldMaxBufferedSpeech = this.#opts.maxBufferedSpeech;\n    this.#opts = { ...this.#opts, ...opts };\n\n    if (this.#inputSampleRate) {\n      // Assert speech buffer exists\n      if (this.#speechBuffer === null) throw new Error('speechBuffer is null');\n\n      // Resize speech buffer\n      this.#prefixPaddingSamples = Math.trunc(\n        (this.#opts.prefixPaddingDuration * this.#inputSampleRate) / 1000,\n      );\n      const bufferSize =\n        Math.trunc((this.#opts.maxBufferedSpeech * this.#inputSampleRate) / 1000) +\n        this.#prefixPaddingSamples;\n      const resizedBuffer = new Int16Array(bufferSize);\n      resizedBuffer.set(\n        this.#speechBuffer.subarray(0, Math.min(this.#speechBuffer.length, bufferSize)),\n      );\n      this.#speechBuffer = resizedBuffer;\n\n      // Determine if max has been reached\n      if (this.#opts.maxBufferedSpeech > oldMaxBufferedSpeech) {\n        this.#speechBufferMaxReached = false;\n      }\n    }\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAGA,oBAOO;AACP,sBAAkE;AAGlE,wBAA+C;AAE/C,MAAM,2BAA2B;AAmBjC,MAAM,oBAAgC;AAAA,EACpC,mBAAmB;AAAA,EACnB,oBAAoB;AAAA,EACpB,uBAAuB;AAAA,EACvB,mBAAmB;AAAA;AAAA,EACnB,qBAAqB;AAAA,EACrB,YAAY;AAAA,EACZ,UAAU;AACZ;AAEO,MAAM,YAAY,cAAAA,IAAQ;AAAA,EAC/B;AAAA,EACA;AAAA,EACA;AAAA,EACA,QAAQ;AAAA,EAER,YAAY,SAA2B,MAAkB;AACvD,UAAM,EAAE,gBAAgB,GAAG,CAAC;AAC5B,SAAK,WAAW;AAChB,SAAK,QAAQ;AACb,SAAK,WAAW,CAAC;AAAA,EACnB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUA,cAAc,MAAiC;AAC7C,SAAK,QAAQ,EAAE,GAAG,KAAK,OAAO,GAAG,KAAK;AACtC,eAAW,UAAU,KAAK,UAAU;AAClC,aAAO,cAAc,KAAK,KAAK;AAAA,IACjC;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EA4BA,aAAa,KAAK,OAA4B,CAAC,GAAiB;AAC9D,UAAM,aAAyB,EAAE,GAAG,mBAAmB,GAAG,KAAK;AAC/D,UAAM,UAAU,UAAM,uCAAoB,WAAW,QAAQ;AAC7D,WAAO,IAAI,IAAI,SAAS,UAAU;AAAA,EACpC;AAAA,EAEA,SAAoB;AAClB,UAAM,SAAS,IAAI;AAAA,MACjB;AAAA,MACA,KAAK;AAAA,MACL,IAAI,4BAAU,KAAK,UAAU,KAAK,MAAM,UAAU;AAAA,IACpD;AACA,SAAK,SAAS,KAAK,MAAM;AACzB,WAAO;AAAA,EACT;AACF;AAEO,MAAM,kBAAkB,cAAAC,UAAW;AAAA,EACxC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA,aAAa,IAAI,wBAAU,IAAI;AAAA,EAC/B,sBAAsB;AAAA,EACtB,cAAU,mBAAI;AAAA,EAEd,YAAY,KAAU,MAAkB,OAAkB;AACxD,UAAM,GAAG;AACT,SAAK,QAAQ;AACb,SAAK,SAAS;AACd,SAAK,mBAAmB;AACxB,SAAK,gBAAgB;AACrB,SAAK,0BAA0B;AAC/B,SAAK,wBAAwB;AAE7B,SAAK,QAAQ,IAAI,QAAQ,YAAY;AACnC,UAAI,gBAAgB,IAAI,aAAa,KAAK,OAAO,iBAAiB;AAGlE,UAAI,oBAAoB;AAGxB,UAAI,cAAc;AAClB,UAAI,oBAAoB;AACxB,UAAI,qBAAqB;AACzB,UAAI,mBAAmB;AACvB,UAAI,eAAe;AACnB,UAAI,0BAA0B;AAC9B,UAAI,2BAA2B;AAE/B,UAAI,cAAc,CAAC;AACnB,UAAI,kBAAgC,CAAC;AACrC,UAAI,YAAmC;AAGvC,UAAI,yBAAyB;AAE7B,uBAAiB,SAAS,KAAK,OAAO;AACpC,YAAI,OAAO,UAAU,UAAU;AAC7B;AAAA,QACF;AAEA,YAAI,CAAC,KAAK,oBAAoB,CAAC,KAAK,eAAe;AACjD,eAAK,mBAAmB,MAAM;AAC9B,eAAK,wBAAwB,KAAK;AAAA,YAC/B,KAAK,MAAM,wBAAwB,KAAK,mBAAoB;AAAA,UAC/D;AACA,gBAAM,aACJ,KAAK,MAAO,KAAK,MAAM,oBAAoB,KAAK,mBAAoB,GAAI,IACxE,KAAK;AACP,eAAK,gBAAgB,IAAI,WAAW,UAAU;AAE9C,cAAI,KAAK,MAAM,eAAe,KAAK,kBAAkB;AAGnD,wBAAY,IAAI;AAAA,cACd,KAAK;AAAA,cACL,KAAK,MAAM;AAAA,cACX;AAAA,cACA,sCAAsB;AAAA;AAAA,YACxB;AAAA,UACF;AAAA,QACF,WAAW,MAAM,eAAe,KAAK,kBAAkB;AACrD,eAAK,QAAQ,MAAM,4DAA4D;AAC/E;AAAA,QACF;AAEA,oBAAY,KAAK,KAAK;AACtB,YAAI,WAAW;AACb,0BAAgB,KAAK,GAAG,UAAU,KAAK,KAAK,CAAC;AAAA,QAC/C,OAAO;AACL,0BAAgB,KAAK,KAAK;AAAA,QAC5B;AAEA,eAAO,MAAM;AACX,gBAAM,YAAY,QAAQ,OAAO,OAAO;AACxC,gBAAM,4BAA4B,gBAC/B,IAAI,CAAC,MAAM,EAAE,iBAAiB,EAC9B,OAAO,CAAC,KAAK,MAAM,MAAM,GAAG,CAAC;AAEhC,cAAI,4BAA4B,KAAK,OAAO,mBAAmB;AAC7D;AAAA,UACF;AAEA,gBAAM,iBAAa,2BAAY,WAAW;AAC1C,gBAAM,qBAAiB,2BAAY,eAAe;AAGlD,0BAAgB,aAAa;AAAA,YAC3B,eAAe,KAAK,SAAS,GAAG,KAAK,OAAO,iBAAiB;AAAA,YAC7D,CAAC,MAAM,IAAI;AAAA,UACb;AAEA,gBAAM,IAAI,MAAM,KAAK,OAClB,IAAI,aAAa,EACjB,KAAK,CAAC,SAAS,KAAK,WAAW,MAAM,GAAG,IAAI,CAAC;AAEhD,gBAAM,iBAAkB,KAAK,OAAO,oBAAoB,KAAK,MAAM,aAAc;AACjF,8BAAoB,KAAK,OAAO;AAChC,0BAAgB;AAChB,gBAAM,kBAAkB,KAAK,mBAAmB,KAAK,OAAO;AAC5D,gBAAM,SAAS,KAAK,OAAO,oBAAoB,kBAAkB;AACjE,gBAAM,YAAY,KAAK,MAAM,MAAM;AACnC,mCAAyB,SAAS;AAGlC,gBAAM,iBAAiB,KAAK,cAAc,SAAS;AACnD,gBAAM,eAAe,KAAK,IAAI,WAAW,cAAc;AACvD,cAAI,eAAe,GAAG;AACpB,iBAAK,cAAc,IAAI,WAAW,KAAK,SAAS,GAAG,YAAY,GAAG,iBAAiB;AACnF,iCAAqB;AAAA,UACvB,WAAW,CAAC,KAAK,yBAAyB;AACxC,iBAAK,0BAA0B;AAC/B,iBAAK,QAAQ;AAAA,cACX;AAAA,YACF;AAAA,UACF;AAEA,gBAAM,oBAAoB,QAAQ,QAAQ,OAAO,OAAO,IAAI,aAAa,OAAO,GAAO,CAAC;AACxF,eAAK,sBAAsB,KAAK;AAAA,YAC9B;AAAA,YACA,KAAK,sBAAsB,oBAAoB;AAAA,UACjD;AACA,cAAI,KAAK,sBAAsB,0BAA0B;AACvD,iBAAK,QACF,MAAM,EAAE,OAAO,KAAK,oBAAoB,CAAC,EACzC,KAAK,mCAAmC;AAAA,UAC7C;AAEA,cAAI,aAAa;AACf,iCAAqB;AAAA,UACvB,OAAO;AACL,kCAAsB;AAAA,UACxB;AAEA,eAAK,MAAM,IAAI;AAAA,YACb,MAAM,2BAAa;AAAA,YACnB,cAAc;AAAA,YACd,WAAW;AAAA,YACX,iBAAiB;AAAA,YACjB,gBAAgB;AAAA,YAChB,aAAa;AAAA,YACb;AAAA,YACA,QAAQ;AAAA,cACN,IAAI;AAAA,gBACF,WAAW,KAAK,SAAS,GAAG,SAAS;AAAA,gBACrC,KAAK;AAAA,gBACL;AAAA,gBACA;AAAA,cACF;AAAA,YACF;AAAA,YACA,UAAU;AAAA,YACV,uBAAuB;AAAA,YACvB,sBAAsB;AAAA,UACxB,CAAC;AAED,gBAAM,mBAAmB,MAAM;AAC7B,gBAAI,CAAC,KAAK,cAAe,OAAM,IAAI,MAAM,uBAAuB;AAChE,gBAAI,qBAAqB,KAAK,uBAAuB;AACnD;AAAA,YACF;AAEA,kBAAM,cAAc,KAAK,cAAc;AAAA,cACrC,oBAAoB,KAAK;AAAA,cACzB;AAAA,YACF;AACA,iBAAK,cAAc,IAAI,aAAa,CAAC;AACrC,gCAAoB,KAAK;AACzB,iBAAK,0BAA0B;AAAA,UACjC;AAEA,gBAAM,mBAAmB,MAAkB;AACzC,gBAAI,CAAC,KAAK,cAAe,OAAM,IAAI,MAAM,uBAAuB;AAChE,mBAAO,IAAI;AAAA,cACT,KAAK,cAAc,SAAS,KAAK,uBAAuB,iBAAiB;AAAA,cACzE,KAAK;AAAA,cACL;AAAA,cACA;AAAA,YACF;AAAA,UACF;AAEA,cAAI,IAAI,KAAK,MAAM,qBAAqB;AACtC,uCAA2B;AAC3B,uCAA2B;AAC3B,gBAAI,CAAC,eAAe,2BAA2B,KAAK,MAAM,mBAAmB;AAC3E,4BAAc;AACd,mCAAqB;AACrB,kCAAoB;AAEpB,mBAAK,MAAM,IAAI;AAAA,gBACb,MAAM,2BAAa;AAAA,gBACnB,cAAc;AAAA,gBACd,WAAW;AAAA,gBACX,iBAAiB;AAAA,gBACjB,gBAAgB;AAAA,gBAChB,aAAa;AAAA,gBACb;AAAA,gBACA,QAAQ,CAAC,iBAAiB,CAAC;AAAA,gBAC3B,UAAU;AAAA,gBACV,uBAAuB;AAAA,gBACvB,sBAAsB;AAAA,cACxB,CAAC;AAAA,YACH;AAAA,UACF,OAAO;AACL,wCAA4B;AAC5B,sCAA0B;AAE1B,gBAAI,CAAC,aAAa;AAChB,+BAAiB;AAAA,YACnB;AAEA,gBAAI,eAAe,2BAA2B,KAAK,MAAM,oBAAoB;AAC3E,4BAAc;AACd,kCAAoB;AACpB,mCAAqB;AAErB,mBAAK,MAAM,IAAI;AAAA,gBACb,MAAM,2BAAa;AAAA,gBACnB,cAAc;AAAA,gBACd,WAAW;AAAA,gBACX,iBAAiB;AAAA,gBACjB,gBAAgB;AAAA,gBAChB,aAAa;AAAA,gBACb;AAAA,gBACA,QAAQ,CAAC,iBAAiB,CAAC;AAAA,gBAC3B,UAAU;AAAA,gBACV,uBAAuB;AAAA,gBACvB,sBAAsB;AAAA,cACxB,CAAC;AAED,+BAAiB;AAAA,YACnB;AAAA,UACF;AAEA,wBAAc,CAAC;AACf,4BAAkB,CAAC;AAEnB,cAAI,WAAW,KAAK,SAAS,WAAW;AACtC,kBAAM,OAAO,WAAW,KAAK,SAAS,SAAS;AAC/C,wBAAY;AAAA,cACV,IAAI,2BAAW,MAAM,KAAK,kBAAkB,GAAG,KAAK,MAAM,KAAK,SAAS,CAAC,CAAC;AAAA,YAC5E;AAAA,UACF;AACA,cAAI,eAAe,KAAK,SAAS,KAAK,OAAO,mBAAmB;AAC9D,kBAAM,OAAO,eAAe,KAAK,SAAS,KAAK,OAAO,iBAAiB;AACvE,4BAAgB;AAAA,cACd,IAAI,2BAAW,MAAM,KAAK,MAAM,YAAY,GAAG,KAAK,MAAM,KAAK,SAAS,CAAC,CAAC;AAAA,YAC5E;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAAA,IACF,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,cAAc,MAA2B;AACvC,UAAM,uBAAuB,KAAK,MAAM;AACxC,SAAK,QAAQ,EAAE,GAAG,KAAK,OAAO,GAAG,KAAK;AAEtC,QAAI,KAAK,kBAAkB;AAEzB,UAAI,KAAK,kBAAkB,KAAM,OAAM,IAAI,MAAM,sBAAsB;AAGvE,WAAK,wBAAwB,KAAK;AAAA,QAC/B,KAAK,MAAM,wBAAwB,KAAK,mBAAoB;AAAA,MAC/D;AACA,YAAM,aACJ,KAAK,MAAO,KAAK,MAAM,oBAAoB,KAAK,mBAAoB,GAAI,IACxE,KAAK;AACP,YAAM,gBAAgB,IAAI,WAAW,UAAU;AAC/C,oBAAc;AAAA,QACZ,KAAK,cAAc,SAAS,GAAG,KAAK,IAAI,KAAK,cAAc,QAAQ,UAAU,CAAC;AAAA,MAChF;AACA,WAAK,gBAAgB;AAGrB,UAAI,KAAK,MAAM,oBAAoB,sBAAsB;AACvD,aAAK,0BAA0B;AAAA,MACjC;AAAA,IACF;AAAA,EACF;AACF;","names":["baseVAD","baseStream"]}

package/dist/vad.js

@@ -11,9 +11,10 @@ import { OnnxModel, newInferenceSession } from "./onnx_model.js";
 const SLOW_INFERENCE_THRESHOLD = 200;
 const defaultVADOptions = {
   minSpeechDuration: 50,
-  minSilenceDuration: 250,
+  minSilenceDuration: 550,
   prefixPaddingDuration: 500,
   maxBufferedSpeech: 6e4,
+  // 60 seconds
   activationThreshold: 0.5,
   sampleRate: 16e3,
   forceCPU: true
@@ -168,7 +169,7 @@ class VADStream extends baseStream {
           const toCopyInt = Math.trunc(toCopy);
           inputCopyRemainingFrac = toCopy - toCopyInt;
           const availableSpace = this.#speechBuffer.length - speechBufferIndex;
-          const toCopyBuffer = Math.min(this.#model.windowSizeSamples, availableSpace);
+          const toCopyBuffer = Math.min(toCopyInt, availableSpace);
           if (toCopyBuffer > 0) {
             this.#speechBuffer.set(inputFrame.data.subarray(0, toCopyBuffer), speechBufferIndex);
             speechBufferIndex += toCopyBuffer;
@@ -187,9 +188,9 @@ class VADStream extends baseStream {
             this.#logger.child({ delay: this.#extraInferenceTime }).warn("inference is slower than realtime");
           }
           if (pubSpeaking) {
-            pubSpeechDuration += inferenceDuration;
+            pubSpeechDuration += windowDuration;
           } else {
-            pubSilenceDuration += inferenceDuration;
+            pubSilenceDuration += windowDuration;
           }
           this.queue.put({
             type: VADEventType.INFERENCE_DONE,

package/dist/vad.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/vad.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2024 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport {\n  ExpFilter,\n  VADEventType,\n  VADStream as baseStream,\n  VAD as baseVAD,\n  log,\n  mergeFrames,\n} from '@livekit/agents';\nimport { AudioFrame, AudioResampler, AudioResamplerQuality } from '@livekit/rtc-node';\nimport type { InferenceSession } from 'onnxruntime-node';\nimport type { SampleRate } from './onnx_model.js';\nimport { OnnxModel, newInferenceSession } from './onnx_model.js';\n\nconst SLOW_INFERENCE_THRESHOLD = 200; // late by 200ms\n\nexport interface VADOptions {\n  /** Minimum duration of speech to start a new speech chunk */\n  minSpeechDuration: number;\n  /** At the end of each speech, wait this duration before ending the speech */\n  minSilenceDuration: number;\n  /** Duration of padding to add to the beginning of each speech chunk */\n  prefixPaddingDuration: number;\n  /** Maximum duration of speech to keep in the buffer */\n  maxBufferedSpeech: number;\n  /** Maximum duration of speech to keep in the buffer*/\n  activationThreshold: number;\n  /** Sample rate for the inference (only 8KHz and 16KHz are supported) */\n  sampleRate: SampleRate;\n  /** Force the use of CPU for inference */\n  forceCPU: boolean;\n}\n\nconst defaultVADOptions: VADOptions = {\n  minSpeechDuration: 50,\n  minSilenceDuration: 250,\n  prefixPaddingDuration: 500,\n  maxBufferedSpeech: 60000,\n  activationThreshold: 0.5,\n  sampleRate: 16000,\n  forceCPU: true,\n};\n\nexport class VAD extends baseVAD {\n  #session: InferenceSession;\n  #opts: VADOptions;\n  #streams: VADStream[];\n  label = 'silero.VAD';\n\n  constructor(session: InferenceSession, opts: VADOptions) {\n    super({ updateInterval: 32 });\n    this.#session = session;\n    this.#opts = opts;\n    this.#streams = [];\n  }\n\n  /**\n   * Updates the VAD options with new values.\n   *\n   * @param opts - Partial options object containing the values to update\n   * @remarks\n   * This method will merge the provided options with existing options and update all active streams.\n   * Only the properties specified in opts will be updated, other properties retain their current values.\n   */\n  updateOptions(opts: Partial<VADOptions>): void {\n    this.#opts = { ...this.#opts, ...opts };\n    for (const stream of this.#streams) {\n      stream.updateOptions(this.#opts);\n    }\n  }\n\n  /**\n   * Load and initialize the Silero VAD model.\n   *\n   * This method loads the ONNX model and prepares it for inference. When options are not provided,\n   * sane defaults are used.\n   *\n   * @remarks\n   * This method may take time to load the model into memory.\n   * It is recommended to call this method inside your prewarm mechanism.\n   *\n   * @example\n   * ```ts\n   * export default defineAgent({\n   *   prewarm: async (proc: JobProcess) => {\n   *     proc.userData.vad = await VAD.load();\n   *   },\n   *   entry: async (ctx: JobContext) => {\n   *     const vad = ctx.proc.userData.vad! as VAD;\n   *     // the rest of your agent logic\n   *   },\n   * });\n   * ```\n   *\n   * @param options -\n   * @returns Promise\\<{@link VAD}\\>: An instance of the VAD class ready for streaming.\n   */\n  static async load(opts: Partial<VADOptions> = {}): Promise<VAD> {\n    const mergedOpts: VADOptions = { ...defaultVADOptions, ...opts };\n    const session = await newInferenceSession(mergedOpts.forceCPU);\n    return new VAD(session, mergedOpts);\n  }\n\n  stream(): VADStream {\n    const stream = new VADStream(\n      this,\n      this.#opts,\n      new OnnxModel(this.#session, this.#opts.sampleRate),\n    );\n    this.#streams.push(stream);\n    return stream;\n  }\n}\n\nexport class VADStream extends baseStream {\n  #opts: VADOptions;\n  #model: OnnxModel;\n  #inputSampleRate: number;\n  #speechBuffer: Int16Array | null;\n  #speechBufferMaxReached: boolean;\n  #prefixPaddingSamples: number;\n  #task: Promise<void>;\n  #expFilter = new ExpFilter(0.35);\n  #extraInferenceTime = 0;\n  #logger = log();\n\n  constructor(vad: VAD, opts: VADOptions, model: OnnxModel) {\n    super(vad);\n    this.#opts = opts;\n    this.#model = model;\n    this.#inputSampleRate = 0;\n    this.#speechBuffer = null;\n    this.#speechBufferMaxReached = false;\n    this.#prefixPaddingSamples = 0;\n\n    this.#task = new Promise(async () => {\n      let inferenceData = new Float32Array(this.#model.windowSizeSamples);\n\n      // a copy is exposed to the user in END_OF_SPEECH\n      let speechBufferIndex = 0;\n\n      // \"pub\" means public, these values are exposed to the users through events\n      let pubSpeaking = false;\n      let pubSpeechDuration = 0;\n      let pubSilenceDuration = 0;\n      let pubCurrentSample = 0;\n      let pubTimestamp = 0;\n      let speechThresholdDuration = 0;\n      let silenceThresholdDuration = 0;\n\n      let inputFrames = [];\n      let inferenceFrames: AudioFrame[] = [];\n      let resampler: AudioResampler | null = null;\n\n      // used to avoid drift when the sampleRate ratio is not an integer\n      let inputCopyRemainingFrac = 0.0;\n\n      for await (const frame of this.input) {\n        if (typeof frame === 'symbol') {\n          continue; // ignore flush sentinel for now\n        }\n\n        if (!this.#inputSampleRate || !this.#speechBuffer) {\n          this.#inputSampleRate = frame.sampleRate;\n          this.#prefixPaddingSamples = Math.trunc(\n            (this.#opts.prefixPaddingDuration * this.#inputSampleRate) / 1000,\n          );\n          const bufferSize =\n            Math.trunc((this.#opts.maxBufferedSpeech * this.#inputSampleRate) / 1000) +\n            this.#prefixPaddingSamples;\n          this.#speechBuffer = new Int16Array(bufferSize);\n\n          if (this.#opts.sampleRate !== this.#inputSampleRate) {\n            // resampling needed: the input sample rate isn't the same as the model's\n            // sample rate used for inference\n            resampler = new AudioResampler(\n              this.#inputSampleRate,\n              this.#opts.sampleRate,\n              1,\n              AudioResamplerQuality.QUICK, // VAD doesn't need high quality\n            );\n          }\n        } else if (frame.sampleRate !== this.#inputSampleRate) {\n          this.#logger.error('a frame with a different sample rate was already published');\n          continue;\n        }\n\n        inputFrames.push(frame);\n        if (resampler) {\n          inferenceFrames.push(...resampler.push(frame));\n        } else {\n          inferenceFrames.push(frame);\n        }\n\n        while (true) {\n          const startTime = process.hrtime.bigint();\n          const availableInferenceSamples = inferenceFrames\n            .map((x) => x.samplesPerChannel)\n            .reduce((acc, x) => acc + x, 0);\n\n          if (availableInferenceSamples < this.#model.windowSizeSamples) {\n            break; // not enough samples to run inference\n          }\n\n          const inputFrame = mergeFrames(inputFrames);\n          const inferenceFrame = mergeFrames(inferenceFrames);\n\n          // convert data to f32\n          inferenceData = Float32Array.from(\n            inferenceFrame.data.subarray(0, this.#model.windowSizeSamples),\n            (x) => x / 32767,\n          );\n\n          const p = await this.#model\n            .run(inferenceData)\n            .then((data) => this.#expFilter.apply(1, data));\n\n          const windowDuration = (this.#model.windowSizeSamples / this.#opts.sampleRate) * 1000;\n          pubCurrentSample += this.#model.windowSizeSamples;\n          pubTimestamp += windowDuration;\n          const resamplingRatio = this.#inputSampleRate / this.#model.sampleRate;\n          const toCopy = this.#model.windowSizeSamples * resamplingRatio + inputCopyRemainingFrac;\n          const toCopyInt = Math.trunc(toCopy);\n          inputCopyRemainingFrac = toCopy - toCopyInt;\n\n          // copy the inference window to the speech buffer\n          const availableSpace = this.#speechBuffer.length - speechBufferIndex;\n          const toCopyBuffer = Math.min(this.#model.windowSizeSamples, availableSpace);\n          if (toCopyBuffer > 0) {\n            this.#speechBuffer.set(inputFrame.data.subarray(0, toCopyBuffer), speechBufferIndex);\n            speechBufferIndex += toCopyBuffer;\n          } else if (!this.#speechBufferMaxReached) {\n            this.#speechBufferMaxReached = true;\n            this.#logger.warn(\n              'maxBufferedSpeech reached, ignoring further data for the current speech input',\n            );\n          }\n\n          const inferenceDuration = Number((process.hrtime.bigint() - startTime) / BigInt(1000000));\n          this.#extraInferenceTime = Math.max(\n            0,\n            this.#extraInferenceTime + inferenceDuration - windowDuration,\n          );\n          if (this.#extraInferenceTime > SLOW_INFERENCE_THRESHOLD) {\n            this.#logger\n              .child({ delay: this.#extraInferenceTime })\n              .warn('inference is slower than realtime');\n          }\n\n          if (pubSpeaking) {\n            pubSpeechDuration += inferenceDuration;\n          } else {\n            pubSilenceDuration += inferenceDuration;\n          }\n\n          this.queue.put({\n            type: VADEventType.INFERENCE_DONE,\n            samplesIndex: pubCurrentSample,\n            timestamp: pubTimestamp,\n            silenceDuration: pubSilenceDuration,\n            speechDuration: pubSpeechDuration,\n            probability: p,\n            inferenceDuration,\n            frames: [\n              new AudioFrame(\n                inputFrame.data.subarray(0, toCopyInt),\n                this.#inputSampleRate,\n                1,\n                toCopyInt,\n              ),\n            ],\n            speaking: pubSpeaking,\n            rawAccumulatedSilence: silenceThresholdDuration,\n            rawAccumulatedSpeech: speechThresholdDuration,\n          });\n\n          const resetWriteCursor = () => {\n            if (!this.#speechBuffer) throw new Error('speechBuffer is empty');\n            if (speechBufferIndex <= this.#prefixPaddingSamples) {\n              return;\n            }\n\n            const paddingData = this.#speechBuffer.subarray(\n              speechBufferIndex - this.#prefixPaddingSamples,\n              speechBufferIndex,\n            );\n            this.#speechBuffer.set(paddingData, 0);\n            speechBufferIndex = this.#prefixPaddingSamples;\n            this.#speechBufferMaxReached = false;\n          };\n\n          const copySpeechBuffer = (): AudioFrame => {\n            if (!this.#speechBuffer) throw new Error('speechBuffer is empty');\n            return new AudioFrame(\n              this.#speechBuffer.subarray(this.#prefixPaddingSamples, speechBufferIndex),\n              this.#inputSampleRate,\n              1,\n              speechBufferIndex,\n            );\n          };\n\n          if (p > this.#opts.activationThreshold) {\n            speechThresholdDuration += windowDuration;\n            silenceThresholdDuration = 0;\n            if (!pubSpeaking && speechThresholdDuration >= this.#opts.minSpeechDuration) {\n              pubSpeaking = true;\n              pubSilenceDuration = 0;\n              pubSpeechDuration = speechThresholdDuration;\n\n              this.queue.put({\n                type: VADEventType.START_OF_SPEECH,\n                samplesIndex: pubCurrentSample,\n                timestamp: pubTimestamp,\n                silenceDuration: pubSilenceDuration,\n                speechDuration: pubSpeechDuration,\n                probability: p,\n                inferenceDuration,\n                frames: [copySpeechBuffer()],\n                speaking: pubSpeaking,\n                rawAccumulatedSilence: 0,\n                rawAccumulatedSpeech: 0,\n              });\n            }\n          } else {\n            silenceThresholdDuration += windowDuration;\n            speechThresholdDuration = 0;\n\n            if (!pubSpeaking) {\n              resetWriteCursor();\n            }\n\n            if (pubSpeaking && silenceThresholdDuration > this.#opts.minSilenceDuration) {\n              pubSpeaking = false;\n              pubSpeechDuration = 0;\n              pubSilenceDuration = silenceThresholdDuration;\n\n              this.queue.put({\n                type: VADEventType.END_OF_SPEECH,\n                samplesIndex: pubCurrentSample,\n                timestamp: pubTimestamp,\n                silenceDuration: pubSilenceDuration,\n                speechDuration: pubSpeechDuration,\n                probability: p,\n                inferenceDuration,\n                frames: [copySpeechBuffer()],\n                speaking: pubSpeaking,\n                rawAccumulatedSilence: 0,\n                rawAccumulatedSpeech: 0,\n              });\n\n              resetWriteCursor();\n            }\n          }\n\n          inputFrames = [];\n          inferenceFrames = [];\n\n          if (inputFrame.data.length > toCopyInt) {\n            const data = inputFrame.data.subarray(toCopyInt);\n            inputFrames.push(\n              new AudioFrame(data, this.#inputSampleRate, 1, Math.trunc(data.length / 2)),\n            );\n          }\n          if (inferenceFrame.data.length > this.#model.windowSizeSamples) {\n            const data = inferenceFrame.data.subarray(this.#model.windowSizeSamples);\n            inferenceFrames.push(\n              new AudioFrame(data, this.#opts.sampleRate, 1, Math.trunc(data.length / 2)),\n            );\n          }\n        }\n      }\n    });\n  }\n\n  /**\n   * Update the VAD options\n   *\n   * @param opts - Partial options object containing the values to update\n   * @remarks\n   * This method allows you to update the VAD options after the VAD object has been created\n   */\n  updateOptions(opts: Partial<VADOptions>) {\n    const oldMaxBufferedSpeech = this.#opts.maxBufferedSpeech;\n    this.#opts = { ...this.#opts, ...opts };\n\n    if (this.#inputSampleRate) {\n      // Assert speech buffer exists\n      if (this.#speechBuffer === null) throw new Error('speechBuffer is null');\n\n      // Resize speech buffer\n      this.#prefixPaddingSamples = Math.trunc(\n        (this.#opts.prefixPaddingDuration * this.#inputSampleRate) / 1000,\n      );\n      const bufferSize =\n        Math.trunc((this.#opts.maxBufferedSpeech * this.#inputSampleRate) / 1000) +\n        this.#prefixPaddingSamples;\n      const resizedBuffer = new Int16Array(bufferSize);\n      resizedBuffer.set(\n        this.#speechBuffer.subarray(0, Math.min(this.#speechBuffer.length, bufferSize)),\n      );\n      this.#speechBuffer = resizedBuffer;\n\n      // Determine if max has been reached\n      if (this.#opts.maxBufferedSpeech > oldMaxBufferedSpeech) {\n        this.#speechBufferMaxReached = false;\n      }\n    }\n  }\n}\n"],"mappings":"AAGA;AAAA,EACE;AAAA,EACA;AAAA,EACA,aAAa;AAAA,EACb,OAAO;AAAA,EACP;AAAA,EACA;AAAA,OACK;AACP,SAAS,YAAY,gBAAgB,6BAA6B;AAGlE,SAAS,WAAW,2BAA2B;AAE/C,MAAM,2BAA2B;AAmBjC,MAAM,oBAAgC;AAAA,EACpC,mBAAmB;AAAA,EACnB,oBAAoB;AAAA,EACpB,uBAAuB;AAAA,EACvB,mBAAmB;AAAA,EACnB,qBAAqB;AAAA,EACrB,YAAY;AAAA,EACZ,UAAU;AACZ;AAEO,MAAM,YAAY,QAAQ;AAAA,EAC/B;AAAA,EACA;AAAA,EACA;AAAA,EACA,QAAQ;AAAA,EAER,YAAY,SAA2B,MAAkB;AACvD,UAAM,EAAE,gBAAgB,GAAG,CAAC;AAC5B,SAAK,WAAW;AAChB,SAAK,QAAQ;AACb,SAAK,WAAW,CAAC;AAAA,EACnB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUA,cAAc,MAAiC;AAC7C,SAAK,QAAQ,EAAE,GAAG,KAAK,OAAO,GAAG,KAAK;AACtC,eAAW,UAAU,KAAK,UAAU;AAClC,aAAO,cAAc,KAAK,KAAK;AAAA,IACjC;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EA4BA,aAAa,KAAK,OAA4B,CAAC,GAAiB;AAC9D,UAAM,aAAyB,EAAE,GAAG,mBAAmB,GAAG,KAAK;AAC/D,UAAM,UAAU,MAAM,oBAAoB,WAAW,QAAQ;AAC7D,WAAO,IAAI,IAAI,SAAS,UAAU;AAAA,EACpC;AAAA,EAEA,SAAoB;AAClB,UAAM,SAAS,IAAI;AAAA,MACjB;AAAA,MACA,KAAK;AAAA,MACL,IAAI,UAAU,KAAK,UAAU,KAAK,MAAM,UAAU;AAAA,IACpD;AACA,SAAK,SAAS,KAAK,MAAM;AACzB,WAAO;AAAA,EACT;AACF;AAEO,MAAM,kBAAkB,WAAW;AAAA,EACxC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA,aAAa,IAAI,UAAU,IAAI;AAAA,EAC/B,sBAAsB;AAAA,EACtB,UAAU,IAAI;AAAA,EAEd,YAAY,KAAU,MAAkB,OAAkB;AACxD,UAAM,GAAG;AACT,SAAK,QAAQ;AACb,SAAK,SAAS;AACd,SAAK,mBAAmB;AACxB,SAAK,gBAAgB;AACrB,SAAK,0BAA0B;AAC/B,SAAK,wBAAwB;AAE7B,SAAK,QAAQ,IAAI,QAAQ,YAAY;AACnC,UAAI,gBAAgB,IAAI,aAAa,KAAK,OAAO,iBAAiB;AAGlE,UAAI,oBAAoB;AAGxB,UAAI,cAAc;AAClB,UAAI,oBAAoB;AACxB,UAAI,qBAAqB;AACzB,UAAI,mBAAmB;AACvB,UAAI,eAAe;AACnB,UAAI,0BAA0B;AAC9B,UAAI,2BAA2B;AAE/B,UAAI,cAAc,CAAC;AACnB,UAAI,kBAAgC,CAAC;AACrC,UAAI,YAAmC;AAGvC,UAAI,yBAAyB;AAE7B,uBAAiB,SAAS,KAAK,OAAO;AACpC,YAAI,OAAO,UAAU,UAAU;AAC7B;AAAA,QACF;AAEA,YAAI,CAAC,KAAK,oBAAoB,CAAC,KAAK,eAAe;AACjD,eAAK,mBAAmB,MAAM;AAC9B,eAAK,wBAAwB,KAAK;AAAA,YAC/B,KAAK,MAAM,wBAAwB,KAAK,mBAAoB;AAAA,UAC/D;AACA,gBAAM,aACJ,KAAK,MAAO,KAAK,MAAM,oBAAoB,KAAK,mBAAoB,GAAI,IACxE,KAAK;AACP,eAAK,gBAAgB,IAAI,WAAW,UAAU;AAE9C,cAAI,KAAK,MAAM,eAAe,KAAK,kBAAkB;AAGnD,wBAAY,IAAI;AAAA,cACd,KAAK;AAAA,cACL,KAAK,MAAM;AAAA,cACX;AAAA,cACA,sBAAsB;AAAA;AAAA,YACxB;AAAA,UACF;AAAA,QACF,WAAW,MAAM,eAAe,KAAK,kBAAkB;AACrD,eAAK,QAAQ,MAAM,4DAA4D;AAC/E;AAAA,QACF;AAEA,oBAAY,KAAK,KAAK;AACtB,YAAI,WAAW;AACb,0BAAgB,KAAK,GAAG,UAAU,KAAK,KAAK,CAAC;AAAA,QAC/C,OAAO;AACL,0BAAgB,KAAK,KAAK;AAAA,QAC5B;AAEA,eAAO,MAAM;AACX,gBAAM,YAAY,QAAQ,OAAO,OAAO;AACxC,gBAAM,4BAA4B,gBAC/B,IAAI,CAAC,MAAM,EAAE,iBAAiB,EAC9B,OAAO,CAAC,KAAK,MAAM,MAAM,GAAG,CAAC;AAEhC,cAAI,4BAA4B,KAAK,OAAO,mBAAmB;AAC7D;AAAA,UACF;AAEA,gBAAM,aAAa,YAAY,WAAW;AAC1C,gBAAM,iBAAiB,YAAY,eAAe;AAGlD,0BAAgB,aAAa;AAAA,YAC3B,eAAe,KAAK,SAAS,GAAG,KAAK,OAAO,iBAAiB;AAAA,YAC7D,CAAC,MAAM,IAAI;AAAA,UACb;AAEA,gBAAM,IAAI,MAAM,KAAK,OAClB,IAAI,aAAa,EACjB,KAAK,CAAC,SAAS,KAAK,WAAW,MAAM,GAAG,IAAI,CAAC;AAEhD,gBAAM,iBAAkB,KAAK,OAAO,oBAAoB,KAAK,MAAM,aAAc;AACjF,8BAAoB,KAAK,OAAO;AAChC,0BAAgB;AAChB,gBAAM,kBAAkB,KAAK,mBAAmB,KAAK,OAAO;AAC5D,gBAAM,SAAS,KAAK,OAAO,oBAAoB,kBAAkB;AACjE,gBAAM,YAAY,KAAK,MAAM,MAAM;AACnC,mCAAyB,SAAS;AAGlC,gBAAM,iBAAiB,KAAK,cAAc,SAAS;AACnD,gBAAM,eAAe,KAAK,IAAI,KAAK,OAAO,mBAAmB,cAAc;AAC3E,cAAI,eAAe,GAAG;AACpB,iBAAK,cAAc,IAAI,WAAW,KAAK,SAAS,GAAG,YAAY,GAAG,iBAAiB;AACnF,iCAAqB;AAAA,UACvB,WAAW,CAAC,KAAK,yBAAyB;AACxC,iBAAK,0BAA0B;AAC/B,iBAAK,QAAQ;AAAA,cACX;AAAA,YACF;AAAA,UACF;AAEA,gBAAM,oBAAoB,QAAQ,QAAQ,OAAO,OAAO,IAAI,aAAa,OAAO,GAAO,CAAC;AACxF,eAAK,sBAAsB,KAAK;AAAA,YAC9B;AAAA,YACA,KAAK,sBAAsB,oBAAoB;AAAA,UACjD;AACA,cAAI,KAAK,sBAAsB,0BAA0B;AACvD,iBAAK,QACF,MAAM,EAAE,OAAO,KAAK,oBAAoB,CAAC,EACzC,KAAK,mCAAmC;AAAA,UAC7C;AAEA,cAAI,aAAa;AACf,iCAAqB;AAAA,UACvB,OAAO;AACL,kCAAsB;AAAA,UACxB;AAEA,eAAK,MAAM,IAAI;AAAA,YACb,MAAM,aAAa;AAAA,YACnB,cAAc;AAAA,YACd,WAAW;AAAA,YACX,iBAAiB;AAAA,YACjB,gBAAgB;AAAA,YAChB,aAAa;AAAA,YACb;AAAA,YACA,QAAQ;AAAA,cACN,IAAI;AAAA,gBACF,WAAW,KAAK,SAAS,GAAG,SAAS;AAAA,gBACrC,KAAK;AAAA,gBACL;AAAA,gBACA;AAAA,cACF;AAAA,YACF;AAAA,YACA,UAAU;AAAA,YACV,uBAAuB;AAAA,YACvB,sBAAsB;AAAA,UACxB,CAAC;AAED,gBAAM,mBAAmB,MAAM;AAC7B,gBAAI,CAAC,KAAK,cAAe,OAAM,IAAI,MAAM,uBAAuB;AAChE,gBAAI,qBAAqB,KAAK,uBAAuB;AACnD;AAAA,YACF;AAEA,kBAAM,cAAc,KAAK,cAAc;AAAA,cACrC,oBAAoB,KAAK;AAAA,cACzB;AAAA,YACF;AACA,iBAAK,cAAc,IAAI,aAAa,CAAC;AACrC,gCAAoB,KAAK;AACzB,iBAAK,0BAA0B;AAAA,UACjC;AAEA,gBAAM,mBAAmB,MAAkB;AACzC,gBAAI,CAAC,KAAK,cAAe,OAAM,IAAI,MAAM,uBAAuB;AAChE,mBAAO,IAAI;AAAA,cACT,KAAK,cAAc,SAAS,KAAK,uBAAuB,iBAAiB;AAAA,cACzE,KAAK;AAAA,cACL;AAAA,cACA;AAAA,YACF;AAAA,UACF;AAEA,cAAI,IAAI,KAAK,MAAM,qBAAqB;AACtC,uCAA2B;AAC3B,uCAA2B;AAC3B,gBAAI,CAAC,eAAe,2BAA2B,KAAK,MAAM,mBAAmB;AAC3E,4BAAc;AACd,mCAAqB;AACrB,kCAAoB;AAEpB,mBAAK,MAAM,IAAI;AAAA,gBACb,MAAM,aAAa;AAAA,gBACnB,cAAc;AAAA,gBACd,WAAW;AAAA,gBACX,iBAAiB;AAAA,gBACjB,gBAAgB;AAAA,gBAChB,aAAa;AAAA,gBACb;AAAA,gBACA,QAAQ,CAAC,iBAAiB,CAAC;AAAA,gBAC3B,UAAU;AAAA,gBACV,uBAAuB;AAAA,gBACvB,sBAAsB;AAAA,cACxB,CAAC;AAAA,YACH;AAAA,UACF,OAAO;AACL,wCAA4B;AAC5B,sCAA0B;AAE1B,gBAAI,CAAC,aAAa;AAChB,+BAAiB;AAAA,YACnB;AAEA,gBAAI,eAAe,2BAA2B,KAAK,MAAM,oBAAoB;AAC3E,4BAAc;AACd,kCAAoB;AACpB,mCAAqB;AAErB,mBAAK,MAAM,IAAI;AAAA,gBACb,MAAM,aAAa;AAAA,gBACnB,cAAc;AAAA,gBACd,WAAW;AAAA,gBACX,iBAAiB;AAAA,gBACjB,gBAAgB;AAAA,gBAChB,aAAa;AAAA,gBACb;AAAA,gBACA,QAAQ,CAAC,iBAAiB,CAAC;AAAA,gBAC3B,UAAU;AAAA,gBACV,uBAAuB;AAAA,gBACvB,sBAAsB;AAAA,cACxB,CAAC;AAED,+BAAiB;AAAA,YACnB;AAAA,UACF;AAEA,wBAAc,CAAC;AACf,4BAAkB,CAAC;AAEnB,cAAI,WAAW,KAAK,SAAS,WAAW;AACtC,kBAAM,OAAO,WAAW,KAAK,SAAS,SAAS;AAC/C,wBAAY;AAAA,cACV,IAAI,WAAW,MAAM,KAAK,kBAAkB,GAAG,KAAK,MAAM,KAAK,SAAS,CAAC,CAAC;AAAA,YAC5E;AAAA,UACF;AACA,cAAI,eAAe,KAAK,SAAS,KAAK,OAAO,mBAAmB;AAC9D,kBAAM,OAAO,eAAe,KAAK,SAAS,KAAK,OAAO,iBAAiB;AACvE,4BAAgB;AAAA,cACd,IAAI,WAAW,MAAM,KAAK,MAAM,YAAY,GAAG,KAAK,MAAM,KAAK,SAAS,CAAC,CAAC;AAAA,YAC5E;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAAA,IACF,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,cAAc,MAA2B;AACvC,UAAM,uBAAuB,KAAK,MAAM;AACxC,SAAK,QAAQ,EAAE,GAAG,KAAK,OAAO,GAAG,KAAK;AAEtC,QAAI,KAAK,kBAAkB;AAEzB,UAAI,KAAK,kBAAkB,KAAM,OAAM,IAAI,MAAM,sBAAsB;AAGvE,WAAK,wBAAwB,KAAK;AAAA,QAC/B,KAAK,MAAM,wBAAwB,KAAK,mBAAoB;AAAA,MAC/D;AACA,YAAM,aACJ,KAAK,MAAO,KAAK,MAAM,oBAAoB,KAAK,mBAAoB,GAAI,IACxE,KAAK;AACP,YAAM,gBAAgB,IAAI,WAAW,UAAU;AAC/C,oBAAc;AAAA,QACZ,KAAK,cAAc,SAAS,GAAG,KAAK,IAAI,KAAK,cAAc,QAAQ,UAAU,CAAC;AAAA,MAChF;AACA,WAAK,gBAAgB;AAGrB,UAAI,KAAK,MAAM,oBAAoB,sBAAsB;AACvD,aAAK,0BAA0B;AAAA,MACjC;AAAA,IACF;AAAA,EACF;AACF;","names":[]}
+{"version":3,"sources":["../src/vad.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2024 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport {\n  ExpFilter,\n  VADEventType,\n  VADStream as baseStream,\n  VAD as baseVAD,\n  log,\n  mergeFrames,\n} from '@livekit/agents';\nimport { AudioFrame, AudioResampler, AudioResamplerQuality } from '@livekit/rtc-node';\nimport type { InferenceSession } from 'onnxruntime-node';\nimport type { SampleRate } from './onnx_model.js';\nimport { OnnxModel, newInferenceSession } from './onnx_model.js';\n\nconst SLOW_INFERENCE_THRESHOLD = 200; // late by 200ms\n\nexport interface VADOptions {\n  /** Minimum duration of speech to start a new speech chunk */\n  minSpeechDuration: number;\n  /** At the end of each speech, wait this duration before ending the speech */\n  minSilenceDuration: number;\n  /** Duration of padding to add to the beginning of each speech chunk */\n  prefixPaddingDuration: number;\n  /** Maximum duration of speech to keep in the buffer */\n  maxBufferedSpeech: number;\n  /** Maximum duration of speech to keep in the buffer*/\n  activationThreshold: number;\n  /** Sample rate for the inference (only 8KHz and 16KHz are supported) */\n  sampleRate: SampleRate;\n  /** Force the use of CPU for inference */\n  forceCPU: boolean;\n}\n\nconst defaultVADOptions: VADOptions = {\n  minSpeechDuration: 50,\n  minSilenceDuration: 550,\n  prefixPaddingDuration: 500,\n  maxBufferedSpeech: 60000, // 60 seconds\n  activationThreshold: 0.5,\n  sampleRate: 16000,\n  forceCPU: true,\n};\n\nexport class VAD extends baseVAD {\n  #session: InferenceSession;\n  #opts: VADOptions;\n  #streams: VADStream[];\n  label = 'silero.VAD';\n\n  constructor(session: InferenceSession, opts: VADOptions) {\n    super({ updateInterval: 32 });\n    this.#session = session;\n    this.#opts = opts;\n    this.#streams = [];\n  }\n\n  /**\n   * Updates the VAD options with new values.\n   *\n   * @param opts - Partial options object containing the values to update\n   * @remarks\n   * This method will merge the provided options with existing options and update all active streams.\n   * Only the properties specified in opts will be updated, other properties retain their current values.\n   */\n  updateOptions(opts: Partial<VADOptions>): void {\n    this.#opts = { ...this.#opts, ...opts };\n    for (const stream of this.#streams) {\n      stream.updateOptions(this.#opts);\n    }\n  }\n\n  /**\n   * Load and initialize the Silero VAD model.\n   *\n   * This method loads the ONNX model and prepares it for inference. When options are not provided,\n   * sane defaults are used.\n   *\n   * @remarks\n   * This method may take time to load the model into memory.\n   * It is recommended to call this method inside your prewarm mechanism.\n   *\n   * @example\n   * ```ts\n   * export default defineAgent({\n   *   prewarm: async (proc: JobProcess) => {\n   *     proc.userData.vad = await VAD.load();\n   *   },\n   *   entry: async (ctx: JobContext) => {\n   *     const vad = ctx.proc.userData.vad! as VAD;\n   *     // the rest of your agent logic\n   *   },\n   * });\n   * ```\n   *\n   * @param options -\n   * @returns Promise\\<{@link VAD}\\>: An instance of the VAD class ready for streaming.\n   */\n  static async load(opts: Partial<VADOptions> = {}): Promise<VAD> {\n    const mergedOpts: VADOptions = { ...defaultVADOptions, ...opts };\n    const session = await newInferenceSession(mergedOpts.forceCPU);\n    return new VAD(session, mergedOpts);\n  }\n\n  stream(): VADStream {\n    const stream = new VADStream(\n      this,\n      this.#opts,\n      new OnnxModel(this.#session, this.#opts.sampleRate),\n    );\n    this.#streams.push(stream);\n    return stream;\n  }\n}\n\nexport class VADStream extends baseStream {\n  #opts: VADOptions;\n  #model: OnnxModel;\n  #inputSampleRate: number;\n  #speechBuffer: Int16Array | null;\n  #speechBufferMaxReached: boolean;\n  #prefixPaddingSamples: number;\n  #task: Promise<void>;\n  #expFilter = new ExpFilter(0.35);\n  #extraInferenceTime = 0;\n  #logger = log();\n\n  constructor(vad: VAD, opts: VADOptions, model: OnnxModel) {\n    super(vad);\n    this.#opts = opts;\n    this.#model = model;\n    this.#inputSampleRate = 0;\n    this.#speechBuffer = null;\n    this.#speechBufferMaxReached = false;\n    this.#prefixPaddingSamples = 0;\n\n    this.#task = new Promise(async () => {\n      let inferenceData = new Float32Array(this.#model.windowSizeSamples);\n\n      // a copy is exposed to the user in END_OF_SPEECH\n      let speechBufferIndex = 0;\n\n      // \"pub\" means public, these values are exposed to the users through events\n      let pubSpeaking = false;\n      let pubSpeechDuration = 0;\n      let pubSilenceDuration = 0;\n      let pubCurrentSample = 0;\n      let pubTimestamp = 0;\n      let speechThresholdDuration = 0;\n      let silenceThresholdDuration = 0;\n\n      let inputFrames = [];\n      let inferenceFrames: AudioFrame[] = [];\n      let resampler: AudioResampler | null = null;\n\n      // used to avoid drift when the sampleRate ratio is not an integer\n      let inputCopyRemainingFrac = 0.0;\n\n      for await (const frame of this.input) {\n        if (typeof frame === 'symbol') {\n          continue; // ignore flush sentinel for now\n        }\n\n        if (!this.#inputSampleRate || !this.#speechBuffer) {\n          this.#inputSampleRate = frame.sampleRate;\n          this.#prefixPaddingSamples = Math.trunc(\n            (this.#opts.prefixPaddingDuration * this.#inputSampleRate) / 1000,\n          );\n          const bufferSize =\n            Math.trunc((this.#opts.maxBufferedSpeech * this.#inputSampleRate) / 1000) +\n            this.#prefixPaddingSamples;\n          this.#speechBuffer = new Int16Array(bufferSize);\n\n          if (this.#opts.sampleRate !== this.#inputSampleRate) {\n            // resampling needed: the input sample rate isn't the same as the model's\n            // sample rate used for inference\n            resampler = new AudioResampler(\n              this.#inputSampleRate,\n              this.#opts.sampleRate,\n              1,\n              AudioResamplerQuality.QUICK, // VAD doesn't need high quality\n            );\n          }\n        } else if (frame.sampleRate !== this.#inputSampleRate) {\n          this.#logger.error('a frame with a different sample rate was already published');\n          continue;\n        }\n\n        inputFrames.push(frame);\n        if (resampler) {\n          inferenceFrames.push(...resampler.push(frame));\n        } else {\n          inferenceFrames.push(frame);\n        }\n\n        while (true) {\n          const startTime = process.hrtime.bigint();\n          const availableInferenceSamples = inferenceFrames\n            .map((x) => x.samplesPerChannel)\n            .reduce((acc, x) => acc + x, 0);\n\n          if (availableInferenceSamples < this.#model.windowSizeSamples) {\n            break; // not enough samples to run inference\n          }\n\n          const inputFrame = mergeFrames(inputFrames);\n          const inferenceFrame = mergeFrames(inferenceFrames);\n\n          // convert data to f32\n          inferenceData = Float32Array.from(\n            inferenceFrame.data.subarray(0, this.#model.windowSizeSamples),\n            (x) => x / 32767,\n          );\n\n          const p = await this.#model\n            .run(inferenceData)\n            .then((data) => this.#expFilter.apply(1, data));\n\n          const windowDuration = (this.#model.windowSizeSamples / this.#opts.sampleRate) * 1000;\n          pubCurrentSample += this.#model.windowSizeSamples;\n          pubTimestamp += windowDuration;\n          const resamplingRatio = this.#inputSampleRate / this.#model.sampleRate;\n          const toCopy = this.#model.windowSizeSamples * resamplingRatio + inputCopyRemainingFrac;\n          const toCopyInt = Math.trunc(toCopy);\n          inputCopyRemainingFrac = toCopy - toCopyInt;\n\n          // copy the inference window to the speech buffer\n          const availableSpace = this.#speechBuffer.length - speechBufferIndex;\n          const toCopyBuffer = Math.min(toCopyInt, availableSpace);\n          if (toCopyBuffer > 0) {\n            this.#speechBuffer.set(inputFrame.data.subarray(0, toCopyBuffer), speechBufferIndex);\n            speechBufferIndex += toCopyBuffer;\n          } else if (!this.#speechBufferMaxReached) {\n            this.#speechBufferMaxReached = true;\n            this.#logger.warn(\n              'maxBufferedSpeech reached, ignoring further data for the current speech input',\n            );\n          }\n\n          const inferenceDuration = Number((process.hrtime.bigint() - startTime) / BigInt(1000000));\n          this.#extraInferenceTime = Math.max(\n            0,\n            this.#extraInferenceTime + inferenceDuration - windowDuration,\n          );\n          if (this.#extraInferenceTime > SLOW_INFERENCE_THRESHOLD) {\n            this.#logger\n              .child({ delay: this.#extraInferenceTime })\n              .warn('inference is slower than realtime');\n          }\n\n          if (pubSpeaking) {\n            pubSpeechDuration += windowDuration;\n          } else {\n            pubSilenceDuration += windowDuration;\n          }\n\n          this.queue.put({\n            type: VADEventType.INFERENCE_DONE,\n            samplesIndex: pubCurrentSample,\n            timestamp: pubTimestamp,\n            silenceDuration: pubSilenceDuration,\n            speechDuration: pubSpeechDuration,\n            probability: p,\n            inferenceDuration,\n            frames: [\n              new AudioFrame(\n                inputFrame.data.subarray(0, toCopyInt),\n                this.#inputSampleRate,\n                1,\n                toCopyInt,\n              ),\n            ],\n            speaking: pubSpeaking,\n            rawAccumulatedSilence: silenceThresholdDuration,\n            rawAccumulatedSpeech: speechThresholdDuration,\n          });\n\n          const resetWriteCursor = () => {\n            if (!this.#speechBuffer) throw new Error('speechBuffer is empty');\n            if (speechBufferIndex <= this.#prefixPaddingSamples) {\n              return;\n            }\n\n            const paddingData = this.#speechBuffer.subarray(\n              speechBufferIndex - this.#prefixPaddingSamples,\n              speechBufferIndex,\n            );\n            this.#speechBuffer.set(paddingData, 0);\n            speechBufferIndex = this.#prefixPaddingSamples;\n            this.#speechBufferMaxReached = false;\n          };\n\n          const copySpeechBuffer = (): AudioFrame => {\n            if (!this.#speechBuffer) throw new Error('speechBuffer is empty');\n            return new AudioFrame(\n              this.#speechBuffer.subarray(this.#prefixPaddingSamples, speechBufferIndex),\n              this.#inputSampleRate,\n              1,\n              speechBufferIndex,\n            );\n          };\n\n          if (p > this.#opts.activationThreshold) {\n            speechThresholdDuration += windowDuration;\n            silenceThresholdDuration = 0;\n            if (!pubSpeaking && speechThresholdDuration >= this.#opts.minSpeechDuration) {\n              pubSpeaking = true;\n              pubSilenceDuration = 0;\n              pubSpeechDuration = speechThresholdDuration;\n\n              this.queue.put({\n                type: VADEventType.START_OF_SPEECH,\n                samplesIndex: pubCurrentSample,\n                timestamp: pubTimestamp,\n                silenceDuration: pubSilenceDuration,\n                speechDuration: pubSpeechDuration,\n                probability: p,\n                inferenceDuration,\n                frames: [copySpeechBuffer()],\n                speaking: pubSpeaking,\n                rawAccumulatedSilence: 0,\n                rawAccumulatedSpeech: 0,\n              });\n            }\n          } else {\n            silenceThresholdDuration += windowDuration;\n            speechThresholdDuration = 0;\n\n            if (!pubSpeaking) {\n              resetWriteCursor();\n            }\n\n            if (pubSpeaking && silenceThresholdDuration > this.#opts.minSilenceDuration) {\n              pubSpeaking = false;\n              pubSpeechDuration = 0;\n              pubSilenceDuration = silenceThresholdDuration;\n\n              this.queue.put({\n                type: VADEventType.END_OF_SPEECH,\n                samplesIndex: pubCurrentSample,\n                timestamp: pubTimestamp,\n                silenceDuration: pubSilenceDuration,\n                speechDuration: pubSpeechDuration,\n                probability: p,\n                inferenceDuration,\n                frames: [copySpeechBuffer()],\n                speaking: pubSpeaking,\n                rawAccumulatedSilence: 0,\n                rawAccumulatedSpeech: 0,\n              });\n\n              resetWriteCursor();\n            }\n          }\n\n          inputFrames = [];\n          inferenceFrames = [];\n\n          if (inputFrame.data.length > toCopyInt) {\n            const data = inputFrame.data.subarray(toCopyInt);\n            inputFrames.push(\n              new AudioFrame(data, this.#inputSampleRate, 1, Math.trunc(data.length / 2)),\n            );\n          }\n          if (inferenceFrame.data.length > this.#model.windowSizeSamples) {\n            const data = inferenceFrame.data.subarray(this.#model.windowSizeSamples);\n            inferenceFrames.push(\n              new AudioFrame(data, this.#opts.sampleRate, 1, Math.trunc(data.length / 2)),\n            );\n          }\n        }\n      }\n    });\n  }\n\n  /**\n   * Update the VAD options\n   *\n   * @param opts - Partial options object containing the values to update\n   * @remarks\n   * This method allows you to update the VAD options after the VAD object has been created\n   */\n  updateOptions(opts: Partial<VADOptions>) {\n    const oldMaxBufferedSpeech = this.#opts.maxBufferedSpeech;\n    this.#opts = { ...this.#opts, ...opts };\n\n    if (this.#inputSampleRate) {\n      // Assert speech buffer exists\n      if (this.#speechBuffer === null) throw new Error('speechBuffer is null');\n\n      // Resize speech buffer\n      this.#prefixPaddingSamples = Math.trunc(\n        (this.#opts.prefixPaddingDuration * this.#inputSampleRate) / 1000,\n      );\n      const bufferSize =\n        Math.trunc((this.#opts.maxBufferedSpeech * this.#inputSampleRate) / 1000) +\n        this.#prefixPaddingSamples;\n      const resizedBuffer = new Int16Array(bufferSize);\n      resizedBuffer.set(\n        this.#speechBuffer.subarray(0, Math.min(this.#speechBuffer.length, bufferSize)),\n      );\n      this.#speechBuffer = resizedBuffer;\n\n      // Determine if max has been reached\n      if (this.#opts.maxBufferedSpeech > oldMaxBufferedSpeech) {\n        this.#speechBufferMaxReached = false;\n      }\n    }\n  }\n}\n"],"mappings":"AAGA;AAAA,EACE;AAAA,EACA;AAAA,EACA,aAAa;AAAA,EACb,OAAO;AAAA,EACP;AAAA,EACA;AAAA,OACK;AACP,SAAS,YAAY,gBAAgB,6BAA6B;AAGlE,SAAS,WAAW,2BAA2B;AAE/C,MAAM,2BAA2B;AAmBjC,MAAM,oBAAgC;AAAA,EACpC,mBAAmB;AAAA,EACnB,oBAAoB;AAAA,EACpB,uBAAuB;AAAA,EACvB,mBAAmB;AAAA;AAAA,EACnB,qBAAqB;AAAA,EACrB,YAAY;AAAA,EACZ,UAAU;AACZ;AAEO,MAAM,YAAY,QAAQ;AAAA,EAC/B;AAAA,EACA;AAAA,EACA;AAAA,EACA,QAAQ;AAAA,EAER,YAAY,SAA2B,MAAkB;AACvD,UAAM,EAAE,gBAAgB,GAAG,CAAC;AAC5B,SAAK,WAAW;AAChB,SAAK,QAAQ;AACb,SAAK,WAAW,CAAC;AAAA,EACnB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUA,cAAc,MAAiC;AAC7C,SAAK,QAAQ,EAAE,GAAG,KAAK,OAAO,GAAG,KAAK;AACtC,eAAW,UAAU,KAAK,UAAU;AAClC,aAAO,cAAc,KAAK,KAAK;AAAA,IACjC;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EA4BA,aAAa,KAAK,OAA4B,CAAC,GAAiB;AAC9D,UAAM,aAAyB,EAAE,GAAG,mBAAmB,GAAG,KAAK;AAC/D,UAAM,UAAU,MAAM,oBAAoB,WAAW,QAAQ;AAC7D,WAAO,IAAI,IAAI,SAAS,UAAU;AAAA,EACpC;AAAA,EAEA,SAAoB;AAClB,UAAM,SAAS,IAAI;AAAA,MACjB;AAAA,MACA,KAAK;AAAA,MACL,IAAI,UAAU,KAAK,UAAU,KAAK,MAAM,UAAU;AAAA,IACpD;AACA,SAAK,SAAS,KAAK,MAAM;AACzB,WAAO;AAAA,EACT;AACF;AAEO,MAAM,kBAAkB,WAAW;AAAA,EACxC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA,aAAa,IAAI,UAAU,IAAI;AAAA,EAC/B,sBAAsB;AAAA,EACtB,UAAU,IAAI;AAAA,EAEd,YAAY,KAAU,MAAkB,OAAkB;AACxD,UAAM,GAAG;AACT,SAAK,QAAQ;AACb,SAAK,SAAS;AACd,SAAK,mBAAmB;AACxB,SAAK,gBAAgB;AACrB,SAAK,0BAA0B;AAC/B,SAAK,wBAAwB;AAE7B,SAAK,QAAQ,IAAI,QAAQ,YAAY;AACnC,UAAI,gBAAgB,IAAI,aAAa,KAAK,OAAO,iBAAiB;AAGlE,UAAI,oBAAoB;AAGxB,UAAI,cAAc;AAClB,UAAI,oBAAoB;AACxB,UAAI,qBAAqB;AACzB,UAAI,mBAAmB;AACvB,UAAI,eAAe;AACnB,UAAI,0BAA0B;AAC9B,UAAI,2BAA2B;AAE/B,UAAI,cAAc,CAAC;AACnB,UAAI,kBAAgC,CAAC;AACrC,UAAI,YAAmC;AAGvC,UAAI,yBAAyB;AAE7B,uBAAiB,SAAS,KAAK,OAAO;AACpC,YAAI,OAAO,UAAU,UAAU;AAC7B;AAAA,QACF;AAEA,YAAI,CAAC,KAAK,oBAAoB,CAAC,KAAK,eAAe;AACjD,eAAK,mBAAmB,MAAM;AAC9B,eAAK,wBAAwB,KAAK;AAAA,YAC/B,KAAK,MAAM,wBAAwB,KAAK,mBAAoB;AAAA,UAC/D;AACA,gBAAM,aACJ,KAAK,MAAO,KAAK,MAAM,oBAAoB,KAAK,mBAAoB,GAAI,IACxE,KAAK;AACP,eAAK,gBAAgB,IAAI,WAAW,UAAU;AAE9C,cAAI,KAAK,MAAM,eAAe,KAAK,kBAAkB;AAGnD,wBAAY,IAAI;AAAA,cACd,KAAK;AAAA,cACL,KAAK,MAAM;AAAA,cACX;AAAA,cACA,sBAAsB;AAAA;AAAA,YACxB;AAAA,UACF;AAAA,QACF,WAAW,MAAM,eAAe,KAAK,kBAAkB;AACrD,eAAK,QAAQ,MAAM,4DAA4D;AAC/E;AAAA,QACF;AAEA,oBAAY,KAAK,KAAK;AACtB,YAAI,WAAW;AACb,0BAAgB,KAAK,GAAG,UAAU,KAAK,KAAK,CAAC;AAAA,QAC/C,OAAO;AACL,0BAAgB,KAAK,KAAK;AAAA,QAC5B;AAEA,eAAO,MAAM;AACX,gBAAM,YAAY,QAAQ,OAAO,OAAO;AACxC,gBAAM,4BAA4B,gBAC/B,IAAI,CAAC,MAAM,EAAE,iBAAiB,EAC9B,OAAO,CAAC,KAAK,MAAM,MAAM,GAAG,CAAC;AAEhC,cAAI,4BAA4B,KAAK,OAAO,mBAAmB;AAC7D;AAAA,UACF;AAEA,gBAAM,aAAa,YAAY,WAAW;AAC1C,gBAAM,iBAAiB,YAAY,eAAe;AAGlD,0BAAgB,aAAa;AAAA,YAC3B,eAAe,KAAK,SAAS,GAAG,KAAK,OAAO,iBAAiB;AAAA,YAC7D,CAAC,MAAM,IAAI;AAAA,UACb;AAEA,gBAAM,IAAI,MAAM,KAAK,OAClB,IAAI,aAAa,EACjB,KAAK,CAAC,SAAS,KAAK,WAAW,MAAM,GAAG,IAAI,CAAC;AAEhD,gBAAM,iBAAkB,KAAK,OAAO,oBAAoB,KAAK,MAAM,aAAc;AACjF,8BAAoB,KAAK,OAAO;AAChC,0BAAgB;AAChB,gBAAM,kBAAkB,KAAK,mBAAmB,KAAK,OAAO;AAC5D,gBAAM,SAAS,KAAK,OAAO,oBAAoB,kBAAkB;AACjE,gBAAM,YAAY,KAAK,MAAM,MAAM;AACnC,mCAAyB,SAAS;AAGlC,gBAAM,iBAAiB,KAAK,cAAc,SAAS;AACnD,gBAAM,eAAe,KAAK,IAAI,WAAW,cAAc;AACvD,cAAI,eAAe,GAAG;AACpB,iBAAK,cAAc,IAAI,WAAW,KAAK,SAAS,GAAG,YAAY,GAAG,iBAAiB;AACnF,iCAAqB;AAAA,UACvB,WAAW,CAAC,KAAK,yBAAyB;AACxC,iBAAK,0BAA0B;AAC/B,iBAAK,QAAQ;AAAA,cACX;AAAA,YACF;AAAA,UACF;AAEA,gBAAM,oBAAoB,QAAQ,QAAQ,OAAO,OAAO,IAAI,aAAa,OAAO,GAAO,CAAC;AACxF,eAAK,sBAAsB,KAAK;AAAA,YAC9B;AAAA,YACA,KAAK,sBAAsB,oBAAoB;AAAA,UACjD;AACA,cAAI,KAAK,sBAAsB,0BAA0B;AACvD,iBAAK,QACF,MAAM,EAAE,OAAO,KAAK,oBAAoB,CAAC,EACzC,KAAK,mCAAmC;AAAA,UAC7C;AAEA,cAAI,aAAa;AACf,iCAAqB;AAAA,UACvB,OAAO;AACL,kCAAsB;AAAA,UACxB;AAEA,eAAK,MAAM,IAAI;AAAA,YACb,MAAM,aAAa;AAAA,YACnB,cAAc;AAAA,YACd,WAAW;AAAA,YACX,iBAAiB;AAAA,YACjB,gBAAgB;AAAA,YAChB,aAAa;AAAA,YACb;AAAA,YACA,QAAQ;AAAA,cACN,IAAI;AAAA,gBACF,WAAW,KAAK,SAAS,GAAG,SAAS;AAAA,gBACrC,KAAK;AAAA,gBACL;AAAA,gBACA;AAAA,cACF;AAAA,YACF;AAAA,YACA,UAAU;AAAA,YACV,uBAAuB;AAAA,YACvB,sBAAsB;AAAA,UACxB,CAAC;AAED,gBAAM,mBAAmB,MAAM;AAC7B,gBAAI,CAAC,KAAK,cAAe,OAAM,IAAI,MAAM,uBAAuB;AAChE,gBAAI,qBAAqB,KAAK,uBAAuB;AACnD;AAAA,YACF;AAEA,kBAAM,cAAc,KAAK,cAAc;AAAA,cACrC,oBAAoB,KAAK;AAAA,cACzB;AAAA,YACF;AACA,iBAAK,cAAc,IAAI,aAAa,CAAC;AACrC,gCAAoB,KAAK;AACzB,iBAAK,0BAA0B;AAAA,UACjC;AAEA,gBAAM,mBAAmB,MAAkB;AACzC,gBAAI,CAAC,KAAK,cAAe,OAAM,IAAI,MAAM,uBAAuB;AAChE,mBAAO,IAAI;AAAA,cACT,KAAK,cAAc,SAAS,KAAK,uBAAuB,iBAAiB;AAAA,cACzE,KAAK;AAAA,cACL;AAAA,cACA;AAAA,YACF;AAAA,UACF;AAEA,cAAI,IAAI,KAAK,MAAM,qBAAqB;AACtC,uCAA2B;AAC3B,uCAA2B;AAC3B,gBAAI,CAAC,eAAe,2BAA2B,KAAK,MAAM,mBAAmB;AAC3E,4BAAc;AACd,mCAAqB;AACrB,kCAAoB;AAEpB,mBAAK,MAAM,IAAI;AAAA,gBACb,MAAM,aAAa;AAAA,gBACnB,cAAc;AAAA,gBACd,WAAW;AAAA,gBACX,iBAAiB;AAAA,gBACjB,gBAAgB;AAAA,gBAChB,aAAa;AAAA,gBACb;AAAA,gBACA,QAAQ,CAAC,iBAAiB,CAAC;AAAA,gBAC3B,UAAU;AAAA,gBACV,uBAAuB;AAAA,gBACvB,sBAAsB;AAAA,cACxB,CAAC;AAAA,YACH;AAAA,UACF,OAAO;AACL,wCAA4B;AAC5B,sCAA0B;AAE1B,gBAAI,CAAC,aAAa;AAChB,+BAAiB;AAAA,YACnB;AAEA,gBAAI,eAAe,2BAA2B,KAAK,MAAM,oBAAoB;AAC3E,4BAAc;AACd,kCAAoB;AACpB,mCAAqB;AAErB,mBAAK,MAAM,IAAI;AAAA,gBACb,MAAM,aAAa;AAAA,gBACnB,cAAc;AAAA,gBACd,WAAW;AAAA,gBACX,iBAAiB;AAAA,gBACjB,gBAAgB;AAAA,gBAChB,aAAa;AAAA,gBACb;AAAA,gBACA,QAAQ,CAAC,iBAAiB,CAAC;AAAA,gBAC3B,UAAU;AAAA,gBACV,uBAAuB;AAAA,gBACvB,sBAAsB;AAAA,cACxB,CAAC;AAED,+BAAiB;AAAA,YACnB;AAAA,UACF;AAEA,wBAAc,CAAC;AACf,4BAAkB,CAAC;AAEnB,cAAI,WAAW,KAAK,SAAS,WAAW;AACtC,kBAAM,OAAO,WAAW,KAAK,SAAS,SAAS;AAC/C,wBAAY;AAAA,cACV,IAAI,WAAW,MAAM,KAAK,kBAAkB,GAAG,KAAK,MAAM,KAAK,SAAS,CAAC,CAAC;AAAA,YAC5E;AAAA,UACF;AACA,cAAI,eAAe,KAAK,SAAS,KAAK,OAAO,mBAAmB;AAC9D,kBAAM,OAAO,eAAe,KAAK,SAAS,KAAK,OAAO,iBAAiB;AACvE,4BAAgB;AAAA,cACd,IAAI,WAAW,MAAM,KAAK,MAAM,YAAY,GAAG,KAAK,MAAM,KAAK,SAAS,CAAC,CAAC;AAAA,YAC5E;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAAA,IACF,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,cAAc,MAA2B;AACvC,UAAM,uBAAuB,KAAK,MAAM;AACxC,SAAK,QAAQ,EAAE,GAAG,KAAK,OAAO,GAAG,KAAK;AAEtC,QAAI,KAAK,kBAAkB;AAEzB,UAAI,KAAK,kBAAkB,KAAM,OAAM,IAAI,MAAM,sBAAsB;AAGvE,WAAK,wBAAwB,KAAK;AAAA,QAC/B,KAAK,MAAM,wBAAwB,KAAK,mBAAoB;AAAA,MAC/D;AACA,YAAM,aACJ,KAAK,MAAO,KAAK,MAAM,oBAAoB,KAAK,mBAAoB,GAAI,IACxE,KAAK;AACP,YAAM,gBAAgB,IAAI,WAAW,UAAU;AAC/C,oBAAc;AAAA,QACZ,KAAK,cAAc,SAAS,GAAG,KAAK,IAAI,KAAK,cAAc,QAAQ,UAAU,CAAC;AAAA,MAChF;AACA,WAAK,gBAAgB;AAGrB,UAAI,KAAK,MAAM,oBAAoB,sBAAsB;AACvD,aAAK,0BAA0B;AAAA,MACjC;AAAA,IACF;AAAA,EACF;AACF;","names":[]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@livekit/agents-plugin-silero",
-  "version": "0.5.7",
+  "version": "0.5.9",
   "description": "Silero voice activity detection LiveKit Node Agents",
   "main": "dist/index.js",
   "require": "dist/index.cjs",
@@ -29,17 +29,17 @@
     "@livekit/rtc-node": "^0.13.11",
     "@microsoft/api-extractor": "^7.35.0",
     "@types/ws": "^8.5.10",
-    "onnxruntime-common": "^1.19.2",
+    "onnxruntime-common": ">=1.19.0 <1.22.0",
     "tsup": "^8.3.5",
     "typescript": "^5.0.0"
   },
   "dependencies": {
-    "onnxruntime-node": "^1.19.2",
+    "onnxruntime-node": ">=1.19.0 <1.22.0",
     "ws": "^8.16.0"
   },
   "peerDependencies": {
     "@livekit/rtc-node": "^0.13.11",
-    "@livekit/agents": "^0.7.7x"
+    "@livekit/agents": "^0.7.9x"
   },
   "scripts": {
     "build": "tsup --onSuccess \"pnpm build:types\" && cp src/silero_vad.onnx dist/",

package/src/vad.ts

@@ -35,9 +35,9 @@ export interface VADOptions {
 
 const defaultVADOptions: VADOptions = {
   minSpeechDuration: 50,
-  minSilenceDuration: 250,
+  minSilenceDuration: 550,
   prefixPaddingDuration: 500,
-  maxBufferedSpeech: 60000,
+  maxBufferedSpeech: 60000, // 60 seconds
   activationThreshold: 0.5,
   sampleRate: 16000,
   forceCPU: true,
@@ -227,7 +227,7 @@ export class VADStream extends baseStream {
 
           // copy the inference window to the speech buffer
           const availableSpace = this.#speechBuffer.length - speechBufferIndex;
-          const toCopyBuffer = Math.min(this.#model.windowSizeSamples, availableSpace);
+          const toCopyBuffer = Math.min(toCopyInt, availableSpace);
           if (toCopyBuffer > 0) {
             this.#speechBuffer.set(inputFrame.data.subarray(0, toCopyBuffer), speechBufferIndex);
             speechBufferIndex += toCopyBuffer;
@@ -250,9 +250,9 @@ export class VADStream extends baseStream {
           }
 
           if (pubSpeaking) {
-            pubSpeechDuration += inferenceDuration;
+            pubSpeechDuration += windowDuration;
           } else {
-            pubSilenceDuration += inferenceDuration;
+            pubSilenceDuration += windowDuration;
           }
 
           this.queue.put({