@fugood/node-whisper-wasm 1.0.19

package/README.md

@@ -0,0 +1,40 @@
+# @fugood/node-whisper-wasm
+
+Browser WASM package for `@fugood/whisper.node`.
+
+The package exposes the same high-level context API as the native packages, but
+model and audio file paths are fetched as URLs and copied into the Emscripten
+filesystem before inference.
+
+```js
+const whisper = await WhisperNodeWasm.initWhisper({
+  filePath: 'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin',
+  useGpu: false,
+})
+
+const { promise } = whisper.transcribeFile('https://raw.githubusercontent.com/ggml-org/whisper.cpp/master/samples/jfk.wav', {
+  language: 'en',
+})
+
+console.log(await promise)
+await whisper.release()
+```
+
+The WASM build uses pthreads, so browsers must serve the page with COOP/COEP
+headers and expose `SharedArrayBuffer`. Whisper transcription defaults to up to
+8 threads based on browser hardware concurrency; pass `maxThreads` to override
+it. Browser pages run model loading, transcription, benchmarks, and VAD in a
+dedicated worker by default so the UI thread can keep rendering. Use
+`configureWasm({ worker: false })` only when you explicitly need the old
+in-thread runtime, or pass `workerUrl`, `indexScriptUrl`, and `runtimeScriptUrl`
+when serving the package files from custom URLs. Set `useGpu: true` only with a
+package built using `GGML_WEBGPU=ON` and a browser that exposes `navigator.gpu`.
+VAD currently falls back to CPU in the browser package because the Silero VAD
+graph hits unsupported WebGPU ops.
+
+The default build emits `whisper-node.js` and `whisper-node.wasm`. Use
+`bash scripts/build-wasm.sh --single-file` only when you want the WASM binary
+embedded into `whisper-node.js`. Modern Emscripten embeds the pthread worker
+bootstrap in the main JS file, so a separate `whisper-node.worker.js` is not
+expected. The package worker that keeps UI work off the main thread is
+`worker.js`.

package/index.d.ts

@@ -0,0 +1,183 @@
+export interface NativeContextOptions {
+  filePath: string
+  modelUrl?: string
+  useFlashAttn?: boolean
+  useGpu?: boolean
+  maxModelBytes?: number
+  worker?: boolean
+}
+
+export interface NativeVadContextOptions {
+  filePath: string
+  modelUrl?: string
+  useGpu?: boolean
+  nThreads?: number
+  maxModelBytes?: number
+  worker?: boolean
+}
+
+export interface TranscribeOptions {
+  language?: string
+  translate?: boolean
+  maxThreads?: number
+  maxContext?: number
+  maxLen?: number
+  tokenTimestamps?: boolean
+  tdrzEnable?: boolean
+  wordThold?: number
+  offset?: number
+  duration?: number
+  temperature?: number
+  temperatureInc?: number
+  beamSize?: number
+  bestOf?: number
+  prompt?: string
+  nProcessors?: number
+  onProgress?: (progress: number) => void
+  onNewSegments?: (result: TranscribeNewSegmentsResult) => void
+}
+
+export interface TranscribeNewSegmentsResult {
+  nNew: number
+  totalNNew: number
+  result: string
+  segments: TranscribeResult['segments']
+}
+
+export interface TranscribeResult {
+  language?: string
+  result: string
+  segments: Array<{
+    text: string
+    t0: number
+    t1: number
+  }>
+  isAborted: boolean
+}
+
+export interface VadOptions {
+  threshold?: number
+  minSpeechDurationMs?: number
+  minSilenceDurationMs?: number
+  maxSpeechDurationS?: number
+  speechPadMs?: number
+  samplesOverlap?: number
+}
+
+export interface VadSegment {
+  t0: number
+  t1: number
+}
+
+export interface BenchResult {
+  config: string
+  nThreads: number
+  encodeMs: number
+  decodeMs: number
+  batchdMs: number
+  promptMs: number
+}
+
+export interface WhisperContext {
+  transcribe(
+    filePath: string,
+    options?: TranscribeOptions,
+  ): {
+    stop: () => Promise<void>
+    promise: Promise<TranscribeResult>
+  }
+  transcribeFile(
+    filePath: string,
+    options?: TranscribeOptions,
+  ): {
+    stop: () => Promise<void>
+    promise: Promise<TranscribeResult>
+  }
+  transcribeData(
+    audioData: ArrayBuffer | ArrayBufferView | Float32Array,
+    options?: TranscribeOptions,
+  ): {
+    stop: () => Promise<void>
+    promise: Promise<TranscribeResult>
+  }
+  bench(nThreads: number): Promise<BenchResult>
+  release(): Promise<void>
+  getModelInfo(): object
+}
+
+export interface WhisperVadContext {
+  detectSpeech(filePath: string, options?: VadOptions): Promise<VadSegment[]>
+  detectSpeechFile(filePath: string, options?: VadOptions): Promise<VadSegment[]>
+  detectSpeechData(
+    audioData: ArrayBuffer | ArrayBufferView | Float32Array,
+    options?: VadOptions,
+  ): Promise<VadSegment[]>
+  release(): Promise<void>
+  getModelInfo(): object
+}
+
+export interface Module {
+  WhisperContext: {
+    new (options: NativeContextOptions): Promise<WhisperContext>
+    toggleNativeLog(
+      enable: boolean,
+      callback?: (level: string, text: string) => void,
+    ): void | Promise<void>
+  }
+  WhisperVadContext: {
+    new (options: NativeVadContextOptions): Promise<WhisperVadContext>
+    toggleNativeLog(
+      enable: boolean,
+      callback?: (level: string, text: string) => void,
+    ): void | Promise<void>
+  }
+}
+
+export interface WasmRuntimeOptions {
+  worker?: boolean
+  workerUrl?: string
+  indexScriptUrl?: string
+  scriptUrl?: string
+  runtimeScriptUrl?: string
+  locateFileBaseUrl?: string
+  locateFile?: (path: string, prefix: string) => string
+  mainScriptUrlOrBlob?: string | Blob
+  print?: (text: string) => void
+  printErr?: (text: string) => void
+}
+
+export declare const WhisperContext: Module['WhisperContext']
+export declare const WhisperVadContext: Module['WhisperVadContext']
+export declare const DEFAULT_WASM_MODEL_SIZE_LIMIT_BYTES: number
+
+export declare function configureWasm(options: WasmRuntimeOptions): void
+export declare function loadWasmModule(): Promise<unknown>
+export declare function loadWhisperModule(): Promise<Module>
+export declare function initWhisper(
+  options: NativeContextOptions,
+): Promise<WhisperContext>
+export declare function initWhisperVad(
+  options: NativeVadContextOptions,
+): Promise<WhisperVadContext>
+export declare function toggleNativeLog(
+  enable: boolean,
+  callback?: (level: string, text: string) => void,
+): Promise<void>
+export declare function addNativeLogListener(
+  listener: (level: string, text: string) => void,
+): { remove: () => void }
+
+declare const _default: {
+  WhisperContext: typeof WhisperContext
+  WhisperVadContext: typeof WhisperVadContext
+  configureWasm: typeof configureWasm
+  loadWasmModule: typeof loadWasmModule
+  loadWhisperModule: typeof loadWhisperModule
+  initWhisper: typeof initWhisper
+  initWhisperVad: typeof initWhisperVad
+  toggleNativeLog: typeof toggleNativeLog
+  addNativeLogListener: typeof addNativeLogListener
+  DEFAULT_WASM_MODEL_SIZE_LIMIT_BYTES: typeof DEFAULT_WASM_MODEL_SIZE_LIMIT_BYTES
+}
+
+export default _default