@livekit/agents-plugin-silero 0.5.0 → 0.5.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (8) hide show
  1. package/dist/vad.cjs +13 -6
  2. package/dist/vad.cjs.map +1 -1
  3. package/dist/vad.d.ts +2 -1
  4. package/dist/vad.d.ts.map +1 -1
  5. package/dist/vad.js +13 -6
  6. package/dist/vad.js.map +1 -1
  7. package/package.json +2 -2
  8. package/src/vad.ts +10 -3
package/dist/vad.cjs CHANGED
@@ -38,6 +38,7 @@ const defaultVADOptions = {
38
38
  class VAD extends import_agents.VAD {
39
39
  #session;
40
40
  #opts;
41
+ label = "silero.VAD";
41
42
  constructor(session, opts) {
42
43
  super({ updateInterval: 32 });
43
44
  this.#session = session;
@@ -75,7 +76,7 @@ class VAD extends import_agents.VAD {
75
76
  return new VAD(session, mergedOpts);
76
77
  }
77
78
  stream() {
78
- return new VADStream(this.#opts, new import_onnx_model.OnnxModel(this.#session, this.#opts.sampleRate));
79
+ return new VADStream(this, this.#opts, new import_onnx_model.OnnxModel(this.#session, this.#opts.sampleRate));
79
80
  }
80
81
  }
81
82
  class VADStream extends import_agents.VADStream {
@@ -85,8 +86,8 @@ class VADStream extends import_agents.VADStream {
85
86
  #expFilter = new import_agents.ExpFilter(0.35);
86
87
  #extraInferenceTime = 0;
87
88
  #logger = (0, import_agents.log)();
88
- constructor(opts, model) {
89
- super();
89
+ constructor(vad, opts, model) {
90
+ super(vad);
90
91
  this.#opts = opts;
91
92
  this.#model = model;
92
93
  this.#task = new Promise(async () => {
@@ -193,7 +194,9 @@ class VADStream extends import_agents.VADStream {
193
194
  frames: [
194
195
  new import_rtc_node.AudioFrame(inputFrame.data.subarray(0, toCopyInt), pubSampleRate, 1, toCopyInt)
195
196
  ],
196
- speaking: pubSpeaking
197
+ speaking: pubSpeaking,
198
+ rawAccumulatedSilence: silenceThresholdDuration,
199
+ rawAccumulatedSpeech: speechThresholdDuration
197
200
  });
198
201
  const resetWriteCursor = () => {
199
202
  if (!speechBuffer) throw new Error("speechBuffer is empty");
@@ -233,7 +236,9 @@ class VADStream extends import_agents.VADStream {
233
236
  probability: p,
234
237
  inferenceDuration,
235
238
  frames: [copySpeechBuffer()],
236
- speaking: pubSpeaking
239
+ speaking: pubSpeaking,
240
+ rawAccumulatedSilence: 0,
241
+ rawAccumulatedSpeech: 0
237
242
  });
238
243
  }
239
244
  } else {
@@ -255,7 +260,9 @@ class VADStream extends import_agents.VADStream {
255
260
  probability: p,
256
261
  inferenceDuration,
257
262
  frames: [copySpeechBuffer()],
258
- speaking: pubSpeaking
263
+ speaking: pubSpeaking,
264
+ rawAccumulatedSilence: 0,
265
+ rawAccumulatedSpeech: 0
259
266
  });
260
267
  resetWriteCursor();
261
268
  }
package/dist/vad.cjs.map CHANGED
@@ -1 +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\n constructor(session: InferenceSession, opts: VADOptions) {\n super({ updateInterval: 32 });\n this.#session = session;\n this.#opts = opts;\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 return new VADStream(this.#opts, new OnnxModel(this.#session, this.#opts.sampleRate));\n }\n}\n\nexport class VADStream extends baseStream {\n #opts: VADOptions;\n #model: OnnxModel;\n #task: Promise<void>;\n #expFilter = new ExpFilter(0.35);\n #extraInferenceTime = 0;\n #logger = log();\n\n constructor(opts: VADOptions, model: OnnxModel) {\n super();\n this.#opts = opts;\n this.#model = model;\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 speechBuffer: Int16Array | null = null;\n let speechBufferMaxReached = false;\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 pubSampleRate = 0;\n let pubPrefixPaddingSamples = 0; // size in samples of padding data\n\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 (!pubSampleRate || !speechBuffer) {\n pubSampleRate = frame.sampleRate;\n pubPrefixPaddingSamples = Math.trunc(\n (this.#opts.prefixPaddingDuration * pubSampleRate) / 1000,\n );\n\n speechBuffer = new Int16Array(\n this.#opts.maxBufferedSpeech * pubSampleRate + pubPrefixPaddingSamples,\n );\n\n if (this.#opts.sampleRate !== pubSampleRate) {\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 pubSampleRate,\n this.#opts.sampleRate,\n 1,\n AudioResamplerQuality.QUICK, // VAD doesn't need high quality\n );\n }\n } else if (frame.sampleRate !== pubSampleRate) {\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 = pubSampleRate / 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 = speechBuffer.length - speechBufferIndex;\n const toCopyBuffer = Math.min(this.#model.windowSizeSamples, availableSpace);\n if (toCopyBuffer > 0) {\n speechBuffer.set(inputFrame.data.subarray(0, toCopyBuffer), speechBufferIndex);\n speechBufferIndex += toCopyBuffer;\n } else if (!speechBufferMaxReached) {\n 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(inputFrame.data.subarray(0, toCopyInt), pubSampleRate, 1, toCopyInt),\n ],\n speaking: pubSpeaking,\n });\n\n const resetWriteCursor = () => {\n if (!speechBuffer) throw new Error('speechBuffer is empty');\n if (speechBufferIndex <= pubPrefixPaddingSamples) {\n return;\n }\n\n const paddingData = speechBuffer.subarray(\n speechBufferIndex - pubPrefixPaddingSamples,\n speechBufferIndex,\n );\n speechBuffer.set(paddingData, 0);\n speechBufferIndex = pubPrefixPaddingSamples;\n speechBufferMaxReached = false;\n };\n\n const copySpeechBuffer = (): AudioFrame => {\n if (!speechBuffer) throw new Error('speechBuffer is empty');\n return new AudioFrame(\n speechBuffer.subarray(0, speechBufferIndex),\n pubSampleRate,\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 });\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 });\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(new AudioFrame(data, pubSampleRate, 1, Math.trunc(data.length / 2)));\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"],"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,EAEA,YAAY,SAA2B,MAAkB;AACvD,UAAM,EAAE,gBAAgB,GAAG,CAAC;AAC5B,SAAK,WAAW;AAChB,SAAK,QAAQ;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,WAAO,IAAI,UAAU,KAAK,OAAO,IAAI,4BAAU,KAAK,UAAU,KAAK,MAAM,UAAU,CAAC;AAAA,EACtF;AACF;AAEO,MAAM,kBAAkB,cAAAC,UAAW;AAAA,EACxC;AAAA,EACA;AAAA,EACA;AAAA,EACA,aAAa,IAAI,wBAAU,IAAI;AAAA,EAC/B,sBAAsB;AAAA,EACtB,cAAU,mBAAI;AAAA,EAEd,YAAY,MAAkB,OAAkB;AAC9C,UAAM;AACN,SAAK,QAAQ;AACb,SAAK,SAAS;AAEd,SAAK,QAAQ,IAAI,QAAQ,YAAY;AACnC,UAAI,gBAAgB,IAAI,aAAa,KAAK,OAAO,iBAAiB;AAGlE,UAAI,eAAkC;AACtC,UAAI,yBAAyB;AAC7B,UAAI,oBAAoB;AAGxB,UAAI,cAAc;AAClB,UAAI,oBAAoB;AACxB,UAAI,qBAAqB;AACzB,UAAI,mBAAmB;AACvB,UAAI,eAAe;AACnB,UAAI,gBAAgB;AACpB,UAAI,0BAA0B;AAE9B,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,iBAAiB,CAAC,cAAc;AACnC,0BAAgB,MAAM;AACtB,oCAA0B,KAAK;AAAA,YAC5B,KAAK,MAAM,wBAAwB,gBAAiB;AAAA,UACvD;AAEA,yBAAe,IAAI;AAAA,YACjB,KAAK,MAAM,oBAAoB,gBAAgB;AAAA,UACjD;AAEA,cAAI,KAAK,MAAM,eAAe,eAAe;AAG3C,wBAAY,IAAI;AAAA,cACd;AAAA,cACA,KAAK,MAAM;AAAA,cACX;AAAA,cACA,sCAAsB;AAAA;AAAA,YACxB;AAAA,UACF;AAAA,QACF,WAAW,MAAM,eAAe,eAAe;AAC7C,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,gBAAgB,KAAK,OAAO;AACpD,gBAAM,SAAS,KAAK,OAAO,oBAAoB,kBAAkB;AACjE,gBAAM,YAAY,KAAK,MAAM,MAAM;AACnC,mCAAyB,SAAS;AAGlC,gBAAM,iBAAiB,aAAa,SAAS;AAC7C,gBAAM,eAAe,KAAK,IAAI,KAAK,OAAO,mBAAmB,cAAc;AAC3E,cAAI,eAAe,GAAG;AACpB,yBAAa,IAAI,WAAW,KAAK,SAAS,GAAG,YAAY,GAAG,iBAAiB;AAC7E,iCAAqB;AAAA,UACvB,WAAW,CAAC,wBAAwB;AAClC,qCAAyB;AACzB,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,2BAAW,WAAW,KAAK,SAAS,GAAG,SAAS,GAAG,eAAe,GAAG,SAAS;AAAA,YACpF;AAAA,YACA,UAAU;AAAA,UACZ,CAAC;AAED,gBAAM,mBAAmB,MAAM;AAC7B,gBAAI,CAAC,aAAc,OAAM,IAAI,MAAM,uBAAuB;AAC1D,gBAAI,qBAAqB,yBAAyB;AAChD;AAAA,YACF;AAEA,kBAAM,cAAc,aAAa;AAAA,cAC/B,oBAAoB;AAAA,cACpB;AAAA,YACF;AACA,yBAAa,IAAI,aAAa,CAAC;AAC/B,gCAAoB;AACpB,qCAAyB;AAAA,UAC3B;AAEA,gBAAM,mBAAmB,MAAkB;AACzC,gBAAI,CAAC,aAAc,OAAM,IAAI,MAAM,uBAAuB;AAC1D,mBAAO,IAAI;AAAA,cACT,aAAa,SAAS,GAAG,iBAAiB;AAAA,cAC1C;AAAA,cACA;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,cACZ,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,cACZ,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,KAAK,IAAI,2BAAW,MAAM,eAAe,GAAG,KAAK,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC;AAAA,UACtF;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;AACF;","names":["baseVAD","baseStream"]}
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 label = 'silero.VAD';\n\n constructor(session: InferenceSession, opts: VADOptions) {\n super({ updateInterval: 32 });\n this.#session = session;\n this.#opts = opts;\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 return new VADStream(this, this.#opts, new OnnxModel(this.#session, this.#opts.sampleRate));\n }\n}\n\nexport class VADStream extends baseStream {\n #opts: VADOptions;\n #model: OnnxModel;\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\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 speechBuffer: Int16Array | null = null;\n let speechBufferMaxReached = false;\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 pubSampleRate = 0;\n let pubPrefixPaddingSamples = 0; // size in samples of padding data\n\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 (!pubSampleRate || !speechBuffer) {\n pubSampleRate = frame.sampleRate;\n pubPrefixPaddingSamples = Math.trunc(\n (this.#opts.prefixPaddingDuration * pubSampleRate) / 1000,\n );\n\n speechBuffer = new Int16Array(\n this.#opts.maxBufferedSpeech * pubSampleRate + pubPrefixPaddingSamples,\n );\n\n if (this.#opts.sampleRate !== pubSampleRate) {\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 pubSampleRate,\n this.#opts.sampleRate,\n 1,\n AudioResamplerQuality.QUICK, // VAD doesn't need high quality\n );\n }\n } else if (frame.sampleRate !== pubSampleRate) {\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 = pubSampleRate / 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 = speechBuffer.length - speechBufferIndex;\n const toCopyBuffer = Math.min(this.#model.windowSizeSamples, availableSpace);\n if (toCopyBuffer > 0) {\n speechBuffer.set(inputFrame.data.subarray(0, toCopyBuffer), speechBufferIndex);\n speechBufferIndex += toCopyBuffer;\n } else if (!speechBufferMaxReached) {\n 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(inputFrame.data.subarray(0, toCopyInt), pubSampleRate, 1, toCopyInt),\n ],\n speaking: pubSpeaking,\n rawAccumulatedSilence: silenceThresholdDuration,\n rawAccumulatedSpeech: speechThresholdDuration,\n });\n\n const resetWriteCursor = () => {\n if (!speechBuffer) throw new Error('speechBuffer is empty');\n if (speechBufferIndex <= pubPrefixPaddingSamples) {\n return;\n }\n\n const paddingData = speechBuffer.subarray(\n speechBufferIndex - pubPrefixPaddingSamples,\n speechBufferIndex,\n );\n speechBuffer.set(paddingData, 0);\n speechBufferIndex = pubPrefixPaddingSamples;\n speechBufferMaxReached = false;\n };\n\n const copySpeechBuffer = (): AudioFrame => {\n if (!speechBuffer) throw new Error('speechBuffer is empty');\n return new AudioFrame(\n speechBuffer.subarray(0, speechBufferIndex),\n pubSampleRate,\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(new AudioFrame(data, pubSampleRate, 1, Math.trunc(data.length / 2)));\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"],"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,QAAQ;AAAA,EAER,YAAY,SAA2B,MAAkB;AACvD,UAAM,EAAE,gBAAgB,GAAG,CAAC;AAC5B,SAAK,WAAW;AAChB,SAAK,QAAQ;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,WAAO,IAAI,UAAU,MAAM,KAAK,OAAO,IAAI,4BAAU,KAAK,UAAU,KAAK,MAAM,UAAU,CAAC;AAAA,EAC5F;AACF;AAEO,MAAM,kBAAkB,cAAAC,UAAW;AAAA,EACxC;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;AAEd,SAAK,QAAQ,IAAI,QAAQ,YAAY;AACnC,UAAI,gBAAgB,IAAI,aAAa,KAAK,OAAO,iBAAiB;AAGlE,UAAI,eAAkC;AACtC,UAAI,yBAAyB;AAC7B,UAAI,oBAAoB;AAGxB,UAAI,cAAc;AAClB,UAAI,oBAAoB;AACxB,UAAI,qBAAqB;AACzB,UAAI,mBAAmB;AACvB,UAAI,eAAe;AACnB,UAAI,gBAAgB;AACpB,UAAI,0BAA0B;AAE9B,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,iBAAiB,CAAC,cAAc;AACnC,0BAAgB,MAAM;AACtB,oCAA0B,KAAK;AAAA,YAC5B,KAAK,MAAM,wBAAwB,gBAAiB;AAAA,UACvD;AAEA,yBAAe,IAAI;AAAA,YACjB,KAAK,MAAM,oBAAoB,gBAAgB;AAAA,UACjD;AAEA,cAAI,KAAK,MAAM,eAAe,eAAe;AAG3C,wBAAY,IAAI;AAAA,cACd;AAAA,cACA,KAAK,MAAM;AAAA,cACX;AAAA,cACA,sCAAsB;AAAA;AAAA,YACxB;AAAA,UACF;AAAA,QACF,WAAW,MAAM,eAAe,eAAe;AAC7C,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,gBAAgB,KAAK,OAAO;AACpD,gBAAM,SAAS,KAAK,OAAO,oBAAoB,kBAAkB;AACjE,gBAAM,YAAY,KAAK,MAAM,MAAM;AACnC,mCAAyB,SAAS;AAGlC,gBAAM,iBAAiB,aAAa,SAAS;AAC7C,gBAAM,eAAe,KAAK,IAAI,KAAK,OAAO,mBAAmB,cAAc;AAC3E,cAAI,eAAe,GAAG;AACpB,yBAAa,IAAI,WAAW,KAAK,SAAS,GAAG,YAAY,GAAG,iBAAiB;AAC7E,iCAAqB;AAAA,UACvB,WAAW,CAAC,wBAAwB;AAClC,qCAAyB;AACzB,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,2BAAW,WAAW,KAAK,SAAS,GAAG,SAAS,GAAG,eAAe,GAAG,SAAS;AAAA,YACpF;AAAA,YACA,UAAU;AAAA,YACV,uBAAuB;AAAA,YACvB,sBAAsB;AAAA,UACxB,CAAC;AAED,gBAAM,mBAAmB,MAAM;AAC7B,gBAAI,CAAC,aAAc,OAAM,IAAI,MAAM,uBAAuB;AAC1D,gBAAI,qBAAqB,yBAAyB;AAChD;AAAA,YACF;AAEA,kBAAM,cAAc,aAAa;AAAA,cAC/B,oBAAoB;AAAA,cACpB;AAAA,YACF;AACA,yBAAa,IAAI,aAAa,CAAC;AAC/B,gCAAoB;AACpB,qCAAyB;AAAA,UAC3B;AAEA,gBAAM,mBAAmB,MAAkB;AACzC,gBAAI,CAAC,aAAc,OAAM,IAAI,MAAM,uBAAuB;AAC1D,mBAAO,IAAI;AAAA,cACT,aAAa,SAAS,GAAG,iBAAiB;AAAA,cAC1C;AAAA,cACA;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,KAAK,IAAI,2BAAW,MAAM,eAAe,GAAG,KAAK,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC;AAAA,UACtF;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;AACF;","names":["baseVAD","baseStream"]}
package/dist/vad.d.ts CHANGED
@@ -21,6 +21,7 @@ export interface VADOptions {
21
21
  }
22
22
  export declare class VAD extends baseVAD {
23
23
  #private;
24
+ label: string;
24
25
  constructor(session: InferenceSession, opts: VADOptions);
25
26
  /**
26
27
  * Load and initialize the Silero VAD model.
@@ -53,6 +54,6 @@ export declare class VAD extends baseVAD {
53
54
  }
54
55
  export declare class VADStream extends baseStream {
55
56
  #private;
56
- constructor(opts: VADOptions, model: OnnxModel);
57
+ constructor(vad: VAD, opts: VADOptions, model: OnnxModel);
57
58
  }
58
59
  //# sourceMappingURL=vad.d.ts.map
package/dist/vad.d.ts.map CHANGED
@@ -1 +1 @@
1
- {"version":3,"file":"vad.d.ts","sourceRoot":"","sources":["../src/vad.ts"],"names":[],"mappings":";AAGA,OAAO,EAGL,SAAS,IAAI,UAAU,EACvB,GAAG,IAAI,OAAO,EAGf,MAAM,iBAAiB,CAAC;AAEzB,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAC;AACzD,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAClD,OAAO,EAAE,SAAS,EAAuB,MAAM,iBAAiB,CAAC;AAIjE,MAAM,WAAW,UAAU;IACzB,6DAA6D;IAC7D,iBAAiB,EAAE,MAAM,CAAC;IAC1B,6EAA6E;IAC7E,kBAAkB,EAAE,MAAM,CAAC;IAC3B,uEAAuE;IACvE,qBAAqB,EAAE,MAAM,CAAC;IAC9B,uDAAuD;IACvD,iBAAiB,EAAE,MAAM,CAAC;IAC1B,sDAAsD;IACtD,mBAAmB,EAAE,MAAM,CAAC;IAC5B,wEAAwE;IACxE,UAAU,EAAE,UAAU,CAAC;IACvB,yCAAyC;IACzC,QAAQ,EAAE,OAAO,CAAC;CACnB;AAYD,qBAAa,GAAI,SAAQ,OAAO;;gBAIlB,OAAO,EAAE,gBAAgB,EAAE,IAAI,EAAE,UAAU;IAMvD;;;;;;;;;;;;;;;;;;;;;;;;;OAyBG;WACU,IAAI,CAAC,IAAI,GAAE,OAAO,CAAC,UAAU,CAAM,GAAG,OAAO,CAAC,GAAG,CAAC;IAM/D,MAAM,IAAI,SAAS;CAGpB;AAED,qBAAa,SAAU,SAAQ,UAAU;;gBAQ3B,IAAI,EAAE,UAAU,EAAE,KAAK,EAAE,SAAS;CA2O/C"}
1
+ {"version":3,"file":"vad.d.ts","sourceRoot":"","sources":["../src/vad.ts"],"names":[],"mappings":";AAGA,OAAO,EAGL,SAAS,IAAI,UAAU,EACvB,GAAG,IAAI,OAAO,EAGf,MAAM,iBAAiB,CAAC;AAEzB,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAC;AACzD,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAClD,OAAO,EAAE,SAAS,EAAuB,MAAM,iBAAiB,CAAC;AAIjE,MAAM,WAAW,UAAU;IACzB,6DAA6D;IAC7D,iBAAiB,EAAE,MAAM,CAAC;IAC1B,6EAA6E;IAC7E,kBAAkB,EAAE,MAAM,CAAC;IAC3B,uEAAuE;IACvE,qBAAqB,EAAE,MAAM,CAAC;IAC9B,uDAAuD;IACvD,iBAAiB,EAAE,MAAM,CAAC;IAC1B,sDAAsD;IACtD,mBAAmB,EAAE,MAAM,CAAC;IAC5B,wEAAwE;IACxE,UAAU,EAAE,UAAU,CAAC;IACvB,yCAAyC;IACzC,QAAQ,EAAE,OAAO,CAAC;CACnB;AAYD,qBAAa,GAAI,SAAQ,OAAO;;IAG9B,KAAK,SAAgB;gBAET,OAAO,EAAE,gBAAgB,EAAE,IAAI,EAAE,UAAU;IAMvD;;;;;;;;;;;;;;;;;;;;;;;;;OAyBG;WACU,IAAI,CAAC,IAAI,GAAE,OAAO,CAAC,UAAU,CAAM,GAAG,OAAO,CAAC,GAAG,CAAC;IAM/D,MAAM,IAAI,SAAS;CAGpB;AAED,qBAAa,SAAU,SAAQ,UAAU;;gBAQ3B,GAAG,EAAE,GAAG,EAAE,IAAI,EAAE,UAAU,EAAE,KAAK,EAAE,SAAS;CAiPzD"}
package/dist/vad.js CHANGED
@@ -21,6 +21,7 @@ const defaultVADOptions = {
21
21
  class VAD extends baseVAD {
22
22
  #session;
23
23
  #opts;
24
+ label = "silero.VAD";
24
25
  constructor(session, opts) {
25
26
  super({ updateInterval: 32 });
26
27
  this.#session = session;
@@ -58,7 +59,7 @@ class VAD extends baseVAD {
58
59
  return new VAD(session, mergedOpts);
59
60
  }
60
61
  stream() {
61
- return new VADStream(this.#opts, new OnnxModel(this.#session, this.#opts.sampleRate));
62
+ return new VADStream(this, this.#opts, new OnnxModel(this.#session, this.#opts.sampleRate));
62
63
  }
63
64
  }
64
65
  class VADStream extends baseStream {
@@ -68,8 +69,8 @@ class VADStream extends baseStream {
68
69
  #expFilter = new ExpFilter(0.35);
69
70
  #extraInferenceTime = 0;
70
71
  #logger = log();
71
- constructor(opts, model) {
72
- super();
72
+ constructor(vad, opts, model) {
73
+ super(vad);
73
74
  this.#opts = opts;
74
75
  this.#model = model;
75
76
  this.#task = new Promise(async () => {
@@ -176,7 +177,9 @@ class VADStream extends baseStream {
176
177
  frames: [
177
178
  new AudioFrame(inputFrame.data.subarray(0, toCopyInt), pubSampleRate, 1, toCopyInt)
178
179
  ],
179
- speaking: pubSpeaking
180
+ speaking: pubSpeaking,
181
+ rawAccumulatedSilence: silenceThresholdDuration,
182
+ rawAccumulatedSpeech: speechThresholdDuration
180
183
  });
181
184
  const resetWriteCursor = () => {
182
185
  if (!speechBuffer) throw new Error("speechBuffer is empty");
@@ -216,7 +219,9 @@ class VADStream extends baseStream {
216
219
  probability: p,
217
220
  inferenceDuration,
218
221
  frames: [copySpeechBuffer()],
219
- speaking: pubSpeaking
222
+ speaking: pubSpeaking,
223
+ rawAccumulatedSilence: 0,
224
+ rawAccumulatedSpeech: 0
220
225
  });
221
226
  }
222
227
  } else {
@@ -238,7 +243,9 @@ class VADStream extends baseStream {
238
243
  probability: p,
239
244
  inferenceDuration,
240
245
  frames: [copySpeechBuffer()],
241
- speaking: pubSpeaking
246
+ speaking: pubSpeaking,
247
+ rawAccumulatedSilence: 0,
248
+ rawAccumulatedSpeech: 0
242
249
  });
243
250
  resetWriteCursor();
244
251
  }
package/dist/vad.js.map CHANGED
@@ -1 +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\n constructor(session: InferenceSession, opts: VADOptions) {\n super({ updateInterval: 32 });\n this.#session = session;\n this.#opts = opts;\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 return new VADStream(this.#opts, new OnnxModel(this.#session, this.#opts.sampleRate));\n }\n}\n\nexport class VADStream extends baseStream {\n #opts: VADOptions;\n #model: OnnxModel;\n #task: Promise<void>;\n #expFilter = new ExpFilter(0.35);\n #extraInferenceTime = 0;\n #logger = log();\n\n constructor(opts: VADOptions, model: OnnxModel) {\n super();\n this.#opts = opts;\n this.#model = model;\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 speechBuffer: Int16Array | null = null;\n let speechBufferMaxReached = false;\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 pubSampleRate = 0;\n let pubPrefixPaddingSamples = 0; // size in samples of padding data\n\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 (!pubSampleRate || !speechBuffer) {\n pubSampleRate = frame.sampleRate;\n pubPrefixPaddingSamples = Math.trunc(\n (this.#opts.prefixPaddingDuration * pubSampleRate) / 1000,\n );\n\n speechBuffer = new Int16Array(\n this.#opts.maxBufferedSpeech * pubSampleRate + pubPrefixPaddingSamples,\n );\n\n if (this.#opts.sampleRate !== pubSampleRate) {\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 pubSampleRate,\n this.#opts.sampleRate,\n 1,\n AudioResamplerQuality.QUICK, // VAD doesn't need high quality\n );\n }\n } else if (frame.sampleRate !== pubSampleRate) {\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 = pubSampleRate / 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 = speechBuffer.length - speechBufferIndex;\n const toCopyBuffer = Math.min(this.#model.windowSizeSamples, availableSpace);\n if (toCopyBuffer > 0) {\n speechBuffer.set(inputFrame.data.subarray(0, toCopyBuffer), speechBufferIndex);\n speechBufferIndex += toCopyBuffer;\n } else if (!speechBufferMaxReached) {\n 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(inputFrame.data.subarray(0, toCopyInt), pubSampleRate, 1, toCopyInt),\n ],\n speaking: pubSpeaking,\n });\n\n const resetWriteCursor = () => {\n if (!speechBuffer) throw new Error('speechBuffer is empty');\n if (speechBufferIndex <= pubPrefixPaddingSamples) {\n return;\n }\n\n const paddingData = speechBuffer.subarray(\n speechBufferIndex - pubPrefixPaddingSamples,\n speechBufferIndex,\n );\n speechBuffer.set(paddingData, 0);\n speechBufferIndex = pubPrefixPaddingSamples;\n speechBufferMaxReached = false;\n };\n\n const copySpeechBuffer = (): AudioFrame => {\n if (!speechBuffer) throw new Error('speechBuffer is empty');\n return new AudioFrame(\n speechBuffer.subarray(0, speechBufferIndex),\n pubSampleRate,\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 });\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 });\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(new AudioFrame(data, pubSampleRate, 1, Math.trunc(data.length / 2)));\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"],"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,EAEA,YAAY,SAA2B,MAAkB;AACvD,UAAM,EAAE,gBAAgB,GAAG,CAAC;AAC5B,SAAK,WAAW;AAChB,SAAK,QAAQ;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,WAAO,IAAI,UAAU,KAAK,OAAO,IAAI,UAAU,KAAK,UAAU,KAAK,MAAM,UAAU,CAAC;AAAA,EACtF;AACF;AAEO,MAAM,kBAAkB,WAAW;AAAA,EACxC;AAAA,EACA;AAAA,EACA;AAAA,EACA,aAAa,IAAI,UAAU,IAAI;AAAA,EAC/B,sBAAsB;AAAA,EACtB,UAAU,IAAI;AAAA,EAEd,YAAY,MAAkB,OAAkB;AAC9C,UAAM;AACN,SAAK,QAAQ;AACb,SAAK,SAAS;AAEd,SAAK,QAAQ,IAAI,QAAQ,YAAY;AACnC,UAAI,gBAAgB,IAAI,aAAa,KAAK,OAAO,iBAAiB;AAGlE,UAAI,eAAkC;AACtC,UAAI,yBAAyB;AAC7B,UAAI,oBAAoB;AAGxB,UAAI,cAAc;AAClB,UAAI,oBAAoB;AACxB,UAAI,qBAAqB;AACzB,UAAI,mBAAmB;AACvB,UAAI,eAAe;AACnB,UAAI,gBAAgB;AACpB,UAAI,0BAA0B;AAE9B,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,iBAAiB,CAAC,cAAc;AACnC,0BAAgB,MAAM;AACtB,oCAA0B,KAAK;AAAA,YAC5B,KAAK,MAAM,wBAAwB,gBAAiB;AAAA,UACvD;AAEA,yBAAe,IAAI;AAAA,YACjB,KAAK,MAAM,oBAAoB,gBAAgB;AAAA,UACjD;AAEA,cAAI,KAAK,MAAM,eAAe,eAAe;AAG3C,wBAAY,IAAI;AAAA,cACd;AAAA,cACA,KAAK,MAAM;AAAA,cACX;AAAA,cACA,sBAAsB;AAAA;AAAA,YACxB;AAAA,UACF;AAAA,QACF,WAAW,MAAM,eAAe,eAAe;AAC7C,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,gBAAgB,KAAK,OAAO;AACpD,gBAAM,SAAS,KAAK,OAAO,oBAAoB,kBAAkB;AACjE,gBAAM,YAAY,KAAK,MAAM,MAAM;AACnC,mCAAyB,SAAS;AAGlC,gBAAM,iBAAiB,aAAa,SAAS;AAC7C,gBAAM,eAAe,KAAK,IAAI,KAAK,OAAO,mBAAmB,cAAc;AAC3E,cAAI,eAAe,GAAG;AACpB,yBAAa,IAAI,WAAW,KAAK,SAAS,GAAG,YAAY,GAAG,iBAAiB;AAC7E,iCAAqB;AAAA,UACvB,WAAW,CAAC,wBAAwB;AAClC,qCAAyB;AACzB,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,WAAW,WAAW,KAAK,SAAS,GAAG,SAAS,GAAG,eAAe,GAAG,SAAS;AAAA,YACpF;AAAA,YACA,UAAU;AAAA,UACZ,CAAC;AAED,gBAAM,mBAAmB,MAAM;AAC7B,gBAAI,CAAC,aAAc,OAAM,IAAI,MAAM,uBAAuB;AAC1D,gBAAI,qBAAqB,yBAAyB;AAChD;AAAA,YACF;AAEA,kBAAM,cAAc,aAAa;AAAA,cAC/B,oBAAoB;AAAA,cACpB;AAAA,YACF;AACA,yBAAa,IAAI,aAAa,CAAC;AAC/B,gCAAoB;AACpB,qCAAyB;AAAA,UAC3B;AAEA,gBAAM,mBAAmB,MAAkB;AACzC,gBAAI,CAAC,aAAc,OAAM,IAAI,MAAM,uBAAuB;AAC1D,mBAAO,IAAI;AAAA,cACT,aAAa,SAAS,GAAG,iBAAiB;AAAA,cAC1C;AAAA,cACA;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,cACZ,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,cACZ,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,KAAK,IAAI,WAAW,MAAM,eAAe,GAAG,KAAK,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC;AAAA,UACtF;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;AACF;","names":[]}
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 label = 'silero.VAD';\n\n constructor(session: InferenceSession, opts: VADOptions) {\n super({ updateInterval: 32 });\n this.#session = session;\n this.#opts = opts;\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 return new VADStream(this, this.#opts, new OnnxModel(this.#session, this.#opts.sampleRate));\n }\n}\n\nexport class VADStream extends baseStream {\n #opts: VADOptions;\n #model: OnnxModel;\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\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 speechBuffer: Int16Array | null = null;\n let speechBufferMaxReached = false;\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 pubSampleRate = 0;\n let pubPrefixPaddingSamples = 0; // size in samples of padding data\n\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 (!pubSampleRate || !speechBuffer) {\n pubSampleRate = frame.sampleRate;\n pubPrefixPaddingSamples = Math.trunc(\n (this.#opts.prefixPaddingDuration * pubSampleRate) / 1000,\n );\n\n speechBuffer = new Int16Array(\n this.#opts.maxBufferedSpeech * pubSampleRate + pubPrefixPaddingSamples,\n );\n\n if (this.#opts.sampleRate !== pubSampleRate) {\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 pubSampleRate,\n this.#opts.sampleRate,\n 1,\n AudioResamplerQuality.QUICK, // VAD doesn't need high quality\n );\n }\n } else if (frame.sampleRate !== pubSampleRate) {\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 = pubSampleRate / 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 = speechBuffer.length - speechBufferIndex;\n const toCopyBuffer = Math.min(this.#model.windowSizeSamples, availableSpace);\n if (toCopyBuffer > 0) {\n speechBuffer.set(inputFrame.data.subarray(0, toCopyBuffer), speechBufferIndex);\n speechBufferIndex += toCopyBuffer;\n } else if (!speechBufferMaxReached) {\n 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(inputFrame.data.subarray(0, toCopyInt), pubSampleRate, 1, toCopyInt),\n ],\n speaking: pubSpeaking,\n rawAccumulatedSilence: silenceThresholdDuration,\n rawAccumulatedSpeech: speechThresholdDuration,\n });\n\n const resetWriteCursor = () => {\n if (!speechBuffer) throw new Error('speechBuffer is empty');\n if (speechBufferIndex <= pubPrefixPaddingSamples) {\n return;\n }\n\n const paddingData = speechBuffer.subarray(\n speechBufferIndex - pubPrefixPaddingSamples,\n speechBufferIndex,\n );\n speechBuffer.set(paddingData, 0);\n speechBufferIndex = pubPrefixPaddingSamples;\n speechBufferMaxReached = false;\n };\n\n const copySpeechBuffer = (): AudioFrame => {\n if (!speechBuffer) throw new Error('speechBuffer is empty');\n return new AudioFrame(\n speechBuffer.subarray(0, speechBufferIndex),\n pubSampleRate,\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(new AudioFrame(data, pubSampleRate, 1, Math.trunc(data.length / 2)));\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"],"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,QAAQ;AAAA,EAER,YAAY,SAA2B,MAAkB;AACvD,UAAM,EAAE,gBAAgB,GAAG,CAAC;AAC5B,SAAK,WAAW;AAChB,SAAK,QAAQ;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,WAAO,IAAI,UAAU,MAAM,KAAK,OAAO,IAAI,UAAU,KAAK,UAAU,KAAK,MAAM,UAAU,CAAC;AAAA,EAC5F;AACF;AAEO,MAAM,kBAAkB,WAAW;AAAA,EACxC;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;AAEd,SAAK,QAAQ,IAAI,QAAQ,YAAY;AACnC,UAAI,gBAAgB,IAAI,aAAa,KAAK,OAAO,iBAAiB;AAGlE,UAAI,eAAkC;AACtC,UAAI,yBAAyB;AAC7B,UAAI,oBAAoB;AAGxB,UAAI,cAAc;AAClB,UAAI,oBAAoB;AACxB,UAAI,qBAAqB;AACzB,UAAI,mBAAmB;AACvB,UAAI,eAAe;AACnB,UAAI,gBAAgB;AACpB,UAAI,0BAA0B;AAE9B,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,iBAAiB,CAAC,cAAc;AACnC,0BAAgB,MAAM;AACtB,oCAA0B,KAAK;AAAA,YAC5B,KAAK,MAAM,wBAAwB,gBAAiB;AAAA,UACvD;AAEA,yBAAe,IAAI;AAAA,YACjB,KAAK,MAAM,oBAAoB,gBAAgB;AAAA,UACjD;AAEA,cAAI,KAAK,MAAM,eAAe,eAAe;AAG3C,wBAAY,IAAI;AAAA,cACd;AAAA,cACA,KAAK,MAAM;AAAA,cACX;AAAA,cACA,sBAAsB;AAAA;AAAA,YACxB;AAAA,UACF;AAAA,QACF,WAAW,MAAM,eAAe,eAAe;AAC7C,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,gBAAgB,KAAK,OAAO;AACpD,gBAAM,SAAS,KAAK,OAAO,oBAAoB,kBAAkB;AACjE,gBAAM,YAAY,KAAK,MAAM,MAAM;AACnC,mCAAyB,SAAS;AAGlC,gBAAM,iBAAiB,aAAa,SAAS;AAC7C,gBAAM,eAAe,KAAK,IAAI,KAAK,OAAO,mBAAmB,cAAc;AAC3E,cAAI,eAAe,GAAG;AACpB,yBAAa,IAAI,WAAW,KAAK,SAAS,GAAG,YAAY,GAAG,iBAAiB;AAC7E,iCAAqB;AAAA,UACvB,WAAW,CAAC,wBAAwB;AAClC,qCAAyB;AACzB,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,WAAW,WAAW,KAAK,SAAS,GAAG,SAAS,GAAG,eAAe,GAAG,SAAS;AAAA,YACpF;AAAA,YACA,UAAU;AAAA,YACV,uBAAuB;AAAA,YACvB,sBAAsB;AAAA,UACxB,CAAC;AAED,gBAAM,mBAAmB,MAAM;AAC7B,gBAAI,CAAC,aAAc,OAAM,IAAI,MAAM,uBAAuB;AAC1D,gBAAI,qBAAqB,yBAAyB;AAChD;AAAA,YACF;AAEA,kBAAM,cAAc,aAAa;AAAA,cAC/B,oBAAoB;AAAA,cACpB;AAAA,YACF;AACA,yBAAa,IAAI,aAAa,CAAC;AAC/B,gCAAoB;AACpB,qCAAyB;AAAA,UAC3B;AAEA,gBAAM,mBAAmB,MAAkB;AACzC,gBAAI,CAAC,aAAc,OAAM,IAAI,MAAM,uBAAuB;AAC1D,mBAAO,IAAI;AAAA,cACT,aAAa,SAAS,GAAG,iBAAiB;AAAA,cAC1C;AAAA,cACA;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,KAAK,IAAI,WAAW,MAAM,eAAe,GAAG,KAAK,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC;AAAA,UACtF;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;AACF;","names":[]}
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@livekit/agents-plugin-silero",
3
- "version": "0.5.0",
3
+ "version": "0.5.1",
4
4
  "description": "Silero voice activity detection LiveKit Node Agents",
5
5
  "main": "dist/index.js",
6
6
  "require": "dist/index.cjs",
@@ -36,7 +36,7 @@
36
36
  },
37
37
  "peerDependencies": {
38
38
  "@livekit/rtc-node": "^0.12.1",
39
- "@livekit/agents": "^0.5.0x"
39
+ "@livekit/agents": "^0.6.0x"
40
40
  },
41
41
  "scripts": {
42
42
  "build": "tsup --onSuccess \"tsc --declaration --emitDeclarationOnly\" && cp src/silero_vad.onnx dist/",
package/src/vad.ts CHANGED
@@ -46,6 +46,7 @@ const defaultVADOptions: VADOptions = {
46
46
  export class VAD extends baseVAD {
47
47
  #session: InferenceSession;
48
48
  #opts: VADOptions;
49
+ label = 'silero.VAD';
49
50
 
50
51
  constructor(session: InferenceSession, opts: VADOptions) {
51
52
  super({ updateInterval: 32 });
@@ -86,7 +87,7 @@ export class VAD extends baseVAD {
86
87
  }
87
88
 
88
89
  stream(): VADStream {
89
- return new VADStream(this.#opts, new OnnxModel(this.#session, this.#opts.sampleRate));
90
+ return new VADStream(this, this.#opts, new OnnxModel(this.#session, this.#opts.sampleRate));
90
91
  }
91
92
  }
92
93
 
@@ -98,8 +99,8 @@ export class VADStream extends baseStream {
98
99
  #extraInferenceTime = 0;
99
100
  #logger = log();
100
101
 
101
- constructor(opts: VADOptions, model: OnnxModel) {
102
- super();
102
+ constructor(vad: VAD, opts: VADOptions, model: OnnxModel) {
103
+ super(vad);
103
104
  this.#opts = opts;
104
105
  this.#model = model;
105
106
 
@@ -240,6 +241,8 @@ export class VADStream extends baseStream {
240
241
  new AudioFrame(inputFrame.data.subarray(0, toCopyInt), pubSampleRate, 1, toCopyInt),
241
242
  ],
242
243
  speaking: pubSpeaking,
244
+ rawAccumulatedSilence: silenceThresholdDuration,
245
+ rawAccumulatedSpeech: speechThresholdDuration,
243
246
  });
244
247
 
245
248
  const resetWriteCursor = () => {
@@ -285,6 +288,8 @@ export class VADStream extends baseStream {
285
288
  inferenceDuration,
286
289
  frames: [copySpeechBuffer()],
287
290
  speaking: pubSpeaking,
291
+ rawAccumulatedSilence: 0,
292
+ rawAccumulatedSpeech: 0,
288
293
  });
289
294
  }
290
295
  } else {
@@ -310,6 +315,8 @@ export class VADStream extends baseStream {
310
315
  inferenceDuration,
311
316
  frames: [copySpeechBuffer()],
312
317
  speaking: pubSpeaking,
318
+ rawAccumulatedSilence: 0,
319
+ rawAccumulatedSpeech: 0,
313
320
  });
314
321
 
315
322
  resetWriteCursor();