@guidekit/vad 0.1.0-beta.1

package/dist/index.d.cts

@@ -0,0 +1,122 @@
+declare const VAD_VERSION = "0.1.0";
+/** Silero VAD frame size: 512 samples at 16 kHz = 32 ms per frame. */
+declare const FRAME_SIZE = 512;
+/** Target sample rate for VAD processing. */
+declare const TARGET_SAMPLE_RATE = 16000;
+interface VADOptions {
+    /** Speech probability threshold (0-1). Default: 0.5 */
+    threshold?: number;
+    /** Minimum speech duration in ms to trigger start. Default: 300 */
+    minSpeechDurationMs?: number;
+    /** Silence duration in ms after speech to trigger end. Default: 500 */
+    silenceDurationMs?: number;
+    /** Sample rate. Default: 16000 */
+    sampleRate?: number;
+    /** Enable debug logging. Default: false */
+    debug?: boolean;
+    /** Custom URL for the Silero VAD ONNX model file. */
+    modelUrl?: string;
+}
+interface VADEvent {
+    type: 'speech-start' | 'speech-end' | 'vad-ready';
+    timestamp: number;
+    /** Speech probability (0-1) at the moment of the event. */
+    probability?: number;
+}
+type VADCallback = (event: VADEvent) => void;
+declare class SileroVAD {
+    private readonly _threshold;
+    private readonly _minSpeechDurationMs;
+    private readonly _silenceDurationMs;
+    private readonly _sampleRate;
+    private readonly _debug;
+    private readonly _modelUrl;
+    private _session;
+    private _h;
+    private _c;
+    private _audioContext;
+    private _ownsAudioContext;
+    private _sourceNode;
+    private _workletNode;
+    private _stream;
+    private _frameBuffer;
+    private _frameBufferOffset;
+    private _isReady;
+    private _isSpeaking;
+    private _isStarted;
+    private _isDestroyed;
+    private _consecutiveSpeechFrames;
+    private _consecutiveSilenceFrames;
+    private _frameDurationMs;
+    private _minSpeechFrames;
+    private _silenceFrames;
+    private _isCalibrating;
+    private _calibrationSamples;
+    private _calibrationFramesNeeded;
+    private _calibratedThreshold;
+    private _listeners;
+    private _processingPromise;
+    constructor(options?: VADOptions);
+    /** Load the ONNX model. Uses Cache API for persistence across sessions. */
+    init(): Promise<void>;
+    /**
+     * Process a single audio frame (512 samples at 16 kHz).
+     * Returns the speech probability (0-1).
+     */
+    processFrame(audioData: Float32Array): Promise<number>;
+    /** Start VAD processing on a MediaStream (typically from getUserMedia). */
+    start(stream: MediaStream): void;
+    /** Stop VAD processing and release audio resources (but keep the model). */
+    stop(): void;
+    /** Register a callback for speech-start events. Returns an unsubscribe function. */
+    onSpeechStart(callback: VADCallback): () => void;
+    /** Register a callback for speech-end events. Returns an unsubscribe function. */
+    onSpeechEnd(callback: VADCallback): () => void;
+    /** Register a callback for vad-ready events. Returns an unsubscribe function. */
+    onReady(callback: VADCallback): () => void;
+    /** Whether the ONNX model is loaded and ready. */
+    get isReady(): boolean;
+    /** Whether speech is currently detected. */
+    get isSpeaking(): boolean;
+    /** The MediaStream currently being processed, or null. */
+    get stream(): MediaStream | null;
+    /** Release ONNX model session and all audio resources. */
+    destroy(): Promise<void>;
+    private _on;
+    private _emit;
+    private _setupAudioPipeline;
+    /**
+     * ScriptProcessorNode fallback (works everywhere, including Safari).
+     * We use a buffer size of 4096 which gives ~85 ms of audio at 48 kHz.
+     */
+    private _setupScriptProcessor;
+    /**
+     * Accumulate resampled audio into FRAME_SIZE chunks and process each full frame.
+     */
+    private _feedAudio;
+    /**
+     * Process a single FRAME_SIZE frame: run inference and update speech state.
+     */
+    private _handleFrame;
+    private _finishCalibration;
+    private _teardownAudioPipeline;
+    /** Reset the LSTM hidden and cell states to zeros. */
+    private _resetStates;
+    private _log;
+}
+/**
+ * Create and initialise a SileroVAD instance in one call.
+ *
+ * ```ts
+ * const vad = await createVAD({ debug: true });
+ * vad.onSpeechStart(() => console.log('speaking'));
+ * vad.start(stream);
+ * ```
+ */
+declare function createVAD(options?: VADOptions): Promise<SileroVAD>;
+
+declare global {
+    var webkitAudioContext: typeof AudioContext | undefined;
+}
+
+export { FRAME_SIZE, SileroVAD, TARGET_SAMPLE_RATE, type VADEvent, type VADOptions, VAD_VERSION, createVAD };

package/dist/index.d.ts

@@ -0,0 +1,122 @@
+declare const VAD_VERSION = "0.1.0";
+/** Silero VAD frame size: 512 samples at 16 kHz = 32 ms per frame. */
+declare const FRAME_SIZE = 512;
+/** Target sample rate for VAD processing. */
+declare const TARGET_SAMPLE_RATE = 16000;
+interface VADOptions {
+    /** Speech probability threshold (0-1). Default: 0.5 */
+    threshold?: number;
+    /** Minimum speech duration in ms to trigger start. Default: 300 */
+    minSpeechDurationMs?: number;
+    /** Silence duration in ms after speech to trigger end. Default: 500 */
+    silenceDurationMs?: number;
+    /** Sample rate. Default: 16000 */
+    sampleRate?: number;
+    /** Enable debug logging. Default: false */
+    debug?: boolean;
+    /** Custom URL for the Silero VAD ONNX model file. */
+    modelUrl?: string;
+}
+interface VADEvent {
+    type: 'speech-start' | 'speech-end' | 'vad-ready';
+    timestamp: number;
+    /** Speech probability (0-1) at the moment of the event. */
+    probability?: number;
+}
+type VADCallback = (event: VADEvent) => void;
+declare class SileroVAD {
+    private readonly _threshold;
+    private readonly _minSpeechDurationMs;
+    private readonly _silenceDurationMs;
+    private readonly _sampleRate;
+    private readonly _debug;
+    private readonly _modelUrl;
+    private _session;
+    private _h;
+    private _c;
+    private _audioContext;
+    private _ownsAudioContext;
+    private _sourceNode;
+    private _workletNode;
+    private _stream;
+    private _frameBuffer;
+    private _frameBufferOffset;
+    private _isReady;
+    private _isSpeaking;
+    private _isStarted;
+    private _isDestroyed;
+    private _consecutiveSpeechFrames;
+    private _consecutiveSilenceFrames;
+    private _frameDurationMs;
+    private _minSpeechFrames;
+    private _silenceFrames;
+    private _isCalibrating;
+    private _calibrationSamples;
+    private _calibrationFramesNeeded;
+    private _calibratedThreshold;
+    private _listeners;
+    private _processingPromise;
+    constructor(options?: VADOptions);
+    /** Load the ONNX model. Uses Cache API for persistence across sessions. */
+    init(): Promise<void>;
+    /**
+     * Process a single audio frame (512 samples at 16 kHz).
+     * Returns the speech probability (0-1).
+     */
+    processFrame(audioData: Float32Array): Promise<number>;
+    /** Start VAD processing on a MediaStream (typically from getUserMedia). */
+    start(stream: MediaStream): void;
+    /** Stop VAD processing and release audio resources (but keep the model). */
+    stop(): void;
+    /** Register a callback for speech-start events. Returns an unsubscribe function. */
+    onSpeechStart(callback: VADCallback): () => void;
+    /** Register a callback for speech-end events. Returns an unsubscribe function. */
+    onSpeechEnd(callback: VADCallback): () => void;
+    /** Register a callback for vad-ready events. Returns an unsubscribe function. */
+    onReady(callback: VADCallback): () => void;
+    /** Whether the ONNX model is loaded and ready. */
+    get isReady(): boolean;
+    /** Whether speech is currently detected. */
+    get isSpeaking(): boolean;
+    /** The MediaStream currently being processed, or null. */
+    get stream(): MediaStream | null;
+    /** Release ONNX model session and all audio resources. */
+    destroy(): Promise<void>;
+    private _on;
+    private _emit;
+    private _setupAudioPipeline;
+    /**
+     * ScriptProcessorNode fallback (works everywhere, including Safari).
+     * We use a buffer size of 4096 which gives ~85 ms of audio at 48 kHz.
+     */
+    private _setupScriptProcessor;
+    /**
+     * Accumulate resampled audio into FRAME_SIZE chunks and process each full frame.
+     */
+    private _feedAudio;
+    /**
+     * Process a single FRAME_SIZE frame: run inference and update speech state.
+     */
+    private _handleFrame;
+    private _finishCalibration;
+    private _teardownAudioPipeline;
+    /** Reset the LSTM hidden and cell states to zeros. */
+    private _resetStates;
+    private _log;
+}
+/**
+ * Create and initialise a SileroVAD instance in one call.
+ *
+ * ```ts
+ * const vad = await createVAD({ debug: true });
+ * vad.onSpeechStart(() => console.log('speaking'));
+ * vad.start(stream);
+ * ```
+ */
+declare function createVAD(options?: VADOptions): Promise<SileroVAD>;
+
+declare global {
+    var webkitAudioContext: typeof AudioContext | undefined;
+}
+
+export { FRAME_SIZE, SileroVAD, TARGET_SAMPLE_RATE, type VADEvent, type VADOptions, VAD_VERSION, createVAD };