omnivad 0.2.0

package/dist/index.d.cts

@@ -0,0 +1,169 @@
+/** Result from non-streaming VAD detection */
+interface VADResult {
+    /** Audio duration in seconds */
+    duration: number;
+    /** Array of [start, end] timestamp pairs in seconds */
+    timestamps: [number, number][];
+}
+/** Result from Audio Event Detection (3-class) */
+interface AEDResult {
+    /** Audio duration in seconds */
+    duration: number;
+    /** Events keyed by type ("speech", "singing", "music") with timestamp pairs */
+    events: Record<string, [number, number][]>;
+    /** Detected duration coverage ratio for each event type */
+    ratios: Record<string, number>;
+}
+/** Per-frame result from streaming VAD */
+interface StreamVADFrameResult {
+    /** Raw probability from model output */
+    confidence: number;
+    /** Currently identical to confidence; reserved for future smoothing */
+    smoothedConfidence: number;
+    /** Whether current frame is classified as speech */
+    isSpeech: boolean;
+    /** 1-based frame index of the emitted frame */
+    frameIndex: number;
+    /** True when speech becomes active at this frame */
+    isSpeechStart: boolean;
+    /** True when speech ends on the previous frame */
+    isSpeechEnd: boolean;
+    /** Start frame of the active or just-finished speech segment */
+    speechStartFrame: number;
+    /** End frame of the just-finished speech segment, or 0 if not ending */
+    speechEndFrame: number;
+}
+/** Full-audio streaming-model output */
+interface StreamVADFullResult {
+    /** Per-frame speech probabilities */
+    probabilities: Float32Array;
+    /** Number of emitted frames */
+    numFrames: number;
+    /** Audio duration in seconds */
+    duration: number;
+}
+/** Configuration for non-streaming VAD */
+interface VADConfig {
+    /** Speech probability threshold (default: 0.4) */
+    speechThreshold?: number;
+    /** Smoothing window size in frames (default: 5) */
+    smoothWindowSize?: number;
+    /** Minimum speech segment length in frames (default: 20) */
+    minSpeechFrames?: number;
+    /** Maximum speech segment length in frames before splitting (default: 2000 = 20s) */
+    maxSpeechFrames?: number;
+    /** Minimum silence segment length in frames for state machine (default: 20) */
+    minSilenceFrames?: number;
+    /** Merge silence segments shorter than this (default: 0 = disabled) */
+    mergeSilenceFrames?: number;
+    /** Extend speech segments by this many frames on each side (default: 0) */
+    extendSpeechFrames?: number;
+}
+/** Configuration for Audio Event Detection */
+interface AEDConfig extends VADConfig {
+    /** Singing probability threshold (default: 0.5) */
+    singingThreshold?: number;
+    /** Music probability threshold (default: 0.5) */
+    musicThreshold?: number;
+}
+/** Configuration for streaming VAD */
+interface StreamVADConfig {
+    /** Speech probability threshold (default: 0.5) */
+    speechThreshold?: number;
+}
+
+/**
+ * Non-streaming Voice Activity Detection (WASM/ncnn backend).
+ *
+ * Audio format:
+ *   - Int16Array: raw 16-bit PCM, converted to normalized float internally
+ *   - Float32Array in [-1.0, 1.0]: normalized audio (Web Audio API format)
+ */
+
+declare class OmniVAD {
+    private handle;
+    private config;
+    private constructor();
+    /**
+     * Create a new OmniVAD instance.
+     * Initializes WASM and loads the bundled ncnn model.
+     */
+    static create(options?: VADConfig): Promise<OmniVAD>;
+    /**
+     * Detect speech segments in audio.
+     *
+     * Accepts Int16Array (PCM) or normalized Float32Array in [-1, 1].
+     */
+    detect(audio: Float32Array | Int16Array): VADResult;
+    /** Release native resources. */
+    dispose(): void;
+}
+
+/**
+ * Streaming Voice Activity Detection (WASM/ncnn backend).
+ * Processes audio frame-by-frame (10ms chunks of 160 samples @ 16kHz).
+ */
+
+declare class OmniStreamVAD {
+    private handle;
+    private inSpeech;
+    private speechStartFrame;
+    private constructor();
+    /**
+     * Create a new OmniStreamVAD instance.
+     * Initializes WASM and loads the bundled ncnn model.
+     */
+    static create(options?: StreamVADConfig): Promise<OmniStreamVAD>;
+    /**
+     * Process one frame of audio (160 int16 samples = 10ms @ 16kHz).
+     * Returns null until enough audio is accumulated.
+     */
+    processFrame(pcm160: Int16Array): StreamVADFrameResult | null;
+    /**
+     * Process entire audio at once and return per-frame probabilities.
+     * @param audio - Float32Array in [-1, 1] or Int16Array of 16kHz mono PCM
+     */
+    detectFull(audio: Float32Array | Int16Array): StreamVADFullResult;
+    /** Reset all internal state. */
+    reset(): void;
+    /** Release native resources. */
+    dispose(): void;
+}
+
+/**
+ * Audio Event Detection: speech, singing, music (WASM/ncnn backend).
+ *
+ * Audio format: same as OmniVAD — Int16Array or normalized Float32Array [-1, 1].
+ */
+
+declare class OmniAED {
+    private handle;
+    private config;
+    private constructor();
+    /**
+     * Create a new OmniAED instance.
+     * Initializes WASM and loads the bundled ncnn model.
+     */
+    static create(options?: AEDConfig): Promise<OmniAED>;
+    /**
+     * Detect audio events (speech, singing, music).
+     *
+     * Accepts Int16Array (PCM) or normalized Float32Array in [-1, 1].
+     */
+    detect(audio: Float32Array | Int16Array): AEDResult;
+    /** Release native resources. */
+    dispose(): void;
+}
+
+/**
+ * Low-level WASM binding for omnivad C API.
+ * Loads the Emscripten module and provides typed wrappers.
+ */
+type EmscriptenModule = any;
+/**
+ * Initialize the WASM module. Call once before using any other functions.
+ * Safe to call multiple times (returns cached module).
+ */
+declare function initWasm(wasmLocator?: (filename: string) => string): Promise<EmscriptenModule>;
+
+export { type AEDConfig, type AEDResult, OmniAED as FireRedAED, OmniStreamVAD as FireRedStreamVAD, OmniVAD as FireRedVAD, OmniAED, OmniStreamVAD, OmniVAD, type StreamVADConfig, type StreamVADFrameResult, type StreamVADFullResult, type VADConfig, type VADResult, initWasm };

package/dist/index.d.ts

@@ -0,0 +1,169 @@
+/** Result from non-streaming VAD detection */
+interface VADResult {
+    /** Audio duration in seconds */
+    duration: number;
+    /** Array of [start, end] timestamp pairs in seconds */
+    timestamps: [number, number][];
+}
+/** Result from Audio Event Detection (3-class) */
+interface AEDResult {
+    /** Audio duration in seconds */
+    duration: number;
+    /** Events keyed by type ("speech", "singing", "music") with timestamp pairs */
+    events: Record<string, [number, number][]>;
+    /** Detected duration coverage ratio for each event type */
+    ratios: Record<string, number>;
+}
+/** Per-frame result from streaming VAD */
+interface StreamVADFrameResult {
+    /** Raw probability from model output */
+    confidence: number;
+    /** Currently identical to confidence; reserved for future smoothing */
+    smoothedConfidence: number;
+    /** Whether current frame is classified as speech */
+    isSpeech: boolean;
+    /** 1-based frame index of the emitted frame */
+    frameIndex: number;
+    /** True when speech becomes active at this frame */
+    isSpeechStart: boolean;
+    /** True when speech ends on the previous frame */
+    isSpeechEnd: boolean;
+    /** Start frame of the active or just-finished speech segment */
+    speechStartFrame: number;
+    /** End frame of the just-finished speech segment, or 0 if not ending */
+    speechEndFrame: number;
+}
+/** Full-audio streaming-model output */
+interface StreamVADFullResult {
+    /** Per-frame speech probabilities */
+    probabilities: Float32Array;
+    /** Number of emitted frames */
+    numFrames: number;
+    /** Audio duration in seconds */
+    duration: number;
+}
+/** Configuration for non-streaming VAD */
+interface VADConfig {
+    /** Speech probability threshold (default: 0.4) */
+    speechThreshold?: number;
+    /** Smoothing window size in frames (default: 5) */
+    smoothWindowSize?: number;
+    /** Minimum speech segment length in frames (default: 20) */
+    minSpeechFrames?: number;
+    /** Maximum speech segment length in frames before splitting (default: 2000 = 20s) */
+    maxSpeechFrames?: number;
+    /** Minimum silence segment length in frames for state machine (default: 20) */
+    minSilenceFrames?: number;
+    /** Merge silence segments shorter than this (default: 0 = disabled) */
+    mergeSilenceFrames?: number;
+    /** Extend speech segments by this many frames on each side (default: 0) */
+    extendSpeechFrames?: number;
+}
+/** Configuration for Audio Event Detection */
+interface AEDConfig extends VADConfig {
+    /** Singing probability threshold (default: 0.5) */
+    singingThreshold?: number;
+    /** Music probability threshold (default: 0.5) */
+    musicThreshold?: number;
+}
+/** Configuration for streaming VAD */
+interface StreamVADConfig {
+    /** Speech probability threshold (default: 0.5) */
+    speechThreshold?: number;
+}
+
+/**
+ * Non-streaming Voice Activity Detection (WASM/ncnn backend).
+ *
+ * Audio format:
+ *   - Int16Array: raw 16-bit PCM, converted to normalized float internally
+ *   - Float32Array in [-1.0, 1.0]: normalized audio (Web Audio API format)
+ */
+
+declare class OmniVAD {
+    private handle;
+    private config;
+    private constructor();
+    /**
+     * Create a new OmniVAD instance.
+     * Initializes WASM and loads the bundled ncnn model.
+     */
+    static create(options?: VADConfig): Promise<OmniVAD>;
+    /**
+     * Detect speech segments in audio.
+     *
+     * Accepts Int16Array (PCM) or normalized Float32Array in [-1, 1].
+     */
+    detect(audio: Float32Array | Int16Array): VADResult;
+    /** Release native resources. */
+    dispose(): void;
+}
+
+/**
+ * Streaming Voice Activity Detection (WASM/ncnn backend).
+ * Processes audio frame-by-frame (10ms chunks of 160 samples @ 16kHz).
+ */
+
+declare class OmniStreamVAD {
+    private handle;
+    private inSpeech;
+    private speechStartFrame;
+    private constructor();
+    /**
+     * Create a new OmniStreamVAD instance.
+     * Initializes WASM and loads the bundled ncnn model.
+     */
+    static create(options?: StreamVADConfig): Promise<OmniStreamVAD>;
+    /**
+     * Process one frame of audio (160 int16 samples = 10ms @ 16kHz).
+     * Returns null until enough audio is accumulated.
+     */
+    processFrame(pcm160: Int16Array): StreamVADFrameResult | null;
+    /**
+     * Process entire audio at once and return per-frame probabilities.
+     * @param audio - Float32Array in [-1, 1] or Int16Array of 16kHz mono PCM
+     */
+    detectFull(audio: Float32Array | Int16Array): StreamVADFullResult;
+    /** Reset all internal state. */
+    reset(): void;
+    /** Release native resources. */
+    dispose(): void;
+}
+
+/**
+ * Audio Event Detection: speech, singing, music (WASM/ncnn backend).
+ *
+ * Audio format: same as OmniVAD — Int16Array or normalized Float32Array [-1, 1].
+ */
+
+declare class OmniAED {
+    private handle;
+    private config;
+    private constructor();
+    /**
+     * Create a new OmniAED instance.
+     * Initializes WASM and loads the bundled ncnn model.
+     */
+    static create(options?: AEDConfig): Promise<OmniAED>;
+    /**
+     * Detect audio events (speech, singing, music).
+     *
+     * Accepts Int16Array (PCM) or normalized Float32Array in [-1, 1].
+     */
+    detect(audio: Float32Array | Int16Array): AEDResult;
+    /** Release native resources. */
+    dispose(): void;
+}
+
+/**
+ * Low-level WASM binding for omnivad C API.
+ * Loads the Emscripten module and provides typed wrappers.
+ */
+type EmscriptenModule = any;
+/**
+ * Initialize the WASM module. Call once before using any other functions.
+ * Safe to call multiple times (returns cached module).
+ */
+declare function initWasm(wasmLocator?: (filename: string) => string): Promise<EmscriptenModule>;
+
+export { type AEDConfig, type AEDResult, OmniAED as FireRedAED, OmniStreamVAD as FireRedStreamVAD, OmniVAD as FireRedVAD, OmniAED, OmniStreamVAD, OmniVAD, type StreamVADConfig, type StreamVADFrameResult, type StreamVADFullResult, type VADConfig, type VADResult, initWasm };