Spaces:
Running
Running
File size: 19,612 Bytes
b8cc2bf c783252 b8cc2bf c783252 b8cc2bf c783252 b8cc2bf c783252 b8cc2bf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 | /**
* Keet v3.0 - Token Stream Transcriber
*
* Uses LCSPTFAMerger for overlapping window transcription with
* token-level merging (NeMo LCS + PTFA enhancements).
*/
import type {
TranscriptionResult,
TranscriptionWord
} from './types';
import { ModelManager } from './ModelManager';
import { LCSPTFAMerger } from 'parakeet.js';
export interface TokenStreamConfig {
/** Window duration in seconds (default 5.0) */
windowDuration?: number;
/** Overlap duration in seconds (default 1.5) */
overlapDuration?: number;
/** Sample rate (default 16000) */
sampleRate?: number;
/** Enable debug logging */
debug?: boolean;
/** Min sequence anchor length for LCS (default 3) */
sequenceAnchorLength?: number;
/** Time tolerance for frame alignment (default 0.15s) */
timeTolerance?: number;
/** Decoder frame stride (default 1). Set to 2 to halve decoder steps at cost of coarser timestamps */
frameStride?: number;
/** Request log probabilities per token for merger arbitration (default false).
* Adds per-frame softmax cost. Disable unless using logProb-based merging. */
returnLogProbs?: boolean;
}
export interface TokenStreamCallbacks {
onConfirmedUpdate?: (text: string, tokens: any[]) => void;
onPendingUpdate?: (text: string, tokens: any[]) => void;
onMergeInfo?: (info: { lcsLength: number; anchorValid: boolean; anchorTokens?: string[] }) => void;
onError?: (error: Error) => void;
}
export interface TokenStreamResult {
/** Confirmed (stable) transcript text */
confirmedText: string;
/** Pending (tentative) transcript text */
pendingText: string;
/** Full transcript (confirmed + pending) */
fullText: string;
/** LCS match length from last merge */
lcsLength: number;
/** Whether anchor was validated */
anchorValid: boolean;
/** Total chunks processed */
chunkCount: number;
/** Tokens used for anchor match (optional) */
anchorTokens?: string[];
}
/**
* TokenStreamTranscriber - Windowed streaming transcription with LCS+PTFA merging.
*
* This class implements the v3 streaming architecture:
* - Short overlapping windows (5-7s)
* - Token-level merging using LCSPTFAMerger
* - Frame-aligned verification
* - Vignetting and logProb arbitration
*/
export class TokenStreamTranscriber {
private _modelManager: ModelManager;
private _merger: any = null; // LCSPTFAMerger instance
private _config: Required<TokenStreamConfig>;
private _callbacks: TokenStreamCallbacks;
// State
private _currentTimestamp: number = 0;
private _chunkCount: number = 0;
private _isProcessing: boolean = false;
private _pendingChunk: { audio: Float32Array; startTime?: number } | null = null;
private _tokenizer: any = null;
constructor(
modelManager: ModelManager,
config: TokenStreamConfig = {},
callbacks: TokenStreamCallbacks = {}
) {
this._modelManager = modelManager;
this._config = {
windowDuration: config.windowDuration ?? 5.0,
overlapDuration: config.overlapDuration ?? 1.5,
sampleRate: config.sampleRate ?? 16000,
debug: config.debug ?? false,
sequenceAnchorLength: config.sequenceAnchorLength ?? 3,
timeTolerance: config.timeTolerance ?? 0.15,
frameStride: config.frameStride ?? 1,
returnLogProbs: config.returnLogProbs ?? false,
};
this._callbacks = callbacks;
}
/**
* Initialize the transcriber with LCSPTFAMerger.
* Must be called after model is loaded.
*/
async initialize(): Promise<void> {
const model = this._modelManager.getModel();
if (!model) {
throw new Error('Model not loaded. Call ModelManager.loadModel() first.');
}
// Get tokenizer from model
this._tokenizer = model.tokenizer;
if (!LCSPTFAMerger) {
throw new Error('LCSPTFAMerger not found in parakeet.js. Make sure you have the latest version.');
}
this._merger = new LCSPTFAMerger({
frameTimeStride: model.getFrameTimeStride?.() ?? 0.08,
timeTolerance: this._config.timeTolerance,
sequenceAnchorLength: this._config.sequenceAnchorLength,
});
if (this._config.debug) {
console.log('[TokenStreamTranscriber] Initialized with LCSPTFAMerger');
}
}
/**
* Update configuration.
* Note: Changing window/overlap requires re-connecting to AudioEngine (or reset).
*/
updateConfig(config: TokenStreamConfig): void {
this._config = {
...this._config,
...config,
// Ensure sensible defaults if partial config provided
windowDuration: config.windowDuration ?? this._config.windowDuration,
overlapDuration: config.overlapDuration ?? this._config.overlapDuration,
};
if (this._config.debug) {
console.log('[TokenStreamTranscriber] Config updated:', this._config);
}
}
/**
* Process an audio chunk through the windowed streaming pipeline.
*
* @param audio - Float32Array of audio samples (16kHz mono PCM)
* @param startTime - The absolute start time of this chunk in seconds
* @returns TokenStreamResult with confirmed/pending text
*/
async processChunk(audio: Float32Array, startTime?: number): Promise<TokenStreamResult> {
if (!this._merger) {
throw new Error('TokenStreamTranscriber not initialized. Call initialize() first.');
}
if (this._isProcessing) {
// Queue the latest chunk instead of dropping it entirely.
// Only the most recent pending chunk is kept (older ones are stale).
this._pendingChunk = { audio, startTime };
if (this._config.debug) {
console.log('[TokenStreamTranscriber] Queued chunk (processing busy)');
}
return this._getEmptyResult();
}
this._isProcessing = true;
try {
const result = await this._processChunkInternal(audio, startTime);
// After finishing, check if a newer chunk was queued while we were busy
if (this._pendingChunk) {
const pending = this._pendingChunk;
this._pendingChunk = null;
// Process the queued chunk immediately (non-blocking schedule)
// We don't await here to avoid blocking the caller, but we do
// need to keep _isProcessing true until it completes.
this._isProcessing = false; // Release lock temporarily
// Use queueMicrotask to process next chunk without starving the event loop
queueMicrotask(() => {
this.processChunk(pending.audio, pending.startTime);
});
}
return result;
} catch (error) {
console.error('[TokenStreamTranscriber] Process chunk error:', error);
this._callbacks.onError?.(error as Error);
throw error;
} finally {
this._isProcessing = false;
}
}
/**
* Internal processing logic extracted for queue support.
*/
private async _processChunkInternal(audio: Float32Array, startTime?: number): Promise<TokenStreamResult> {
const model = this._modelManager.getModel();
if (!model) {
throw new Error('Model not available');
}
// Use provided startTime or fall back to internal tracking
const chunkStartTime = startTime !== undefined ? startTime : this._currentTimestamp;
// Calculate actual overlap based on previous chunk end
let actualOverlap = 0;
if (this._chunkCount > 0) {
// If we use explicit startTime, we can calculate overlap precisely:
// overlap = previousWindowEnd - currentWindowStart
const previousWindowEnd = this._currentTimestamp + (audio.length / this._config.sampleRate);
actualOverlap = Math.max(0, previousWindowEnd - chunkStartTime);
}
// Transcribe with all metadata needed for merging.
// Use incremental cache to skip re-decoding the overlap prefix.
// On subsequent chunks, the decoder resumes from cached state at
// the boundary of the overlap, saving ~80% of decode frames.
//
// prefixSamples enables incremental mel caching (JS preprocessor):
// the overlap portion's mel features are reused from the previous call,
// saving ~60-70% of preprocessing time for typical 70% overlap.
const overlapSeconds = this._chunkCount > 0 ? actualOverlap : 0;
const overlapSamples = Math.floor(overlapSeconds * this._config.sampleRate);
const result = await model.transcribe(audio, this._config.sampleRate, {
returnTimestamps: true,
returnTokenIds: true,
returnFrameIndices: true,
returnLogProbs: this._config.returnLogProbs,
timeOffset: chunkStartTime,
frameStride: this._config.frameStride,
prefixSamples: overlapSamples > 0 ? overlapSamples : 0,
incremental: overlapSeconds > 0 ? {
cacheKey: 'streaming',
prefixSeconds: overlapSeconds,
} : undefined,
});
// Merge using LCSPTFAMerger
// We use the provided overlap or calculated one
const mergeResult = this._merger!.processChunk(
result,
chunkStartTime,
this._chunkCount > 0 ? (startTime !== undefined ? actualOverlap : this._config.overlapDuration) : 0
);
// Update state for next chunk
this._currentTimestamp = chunkStartTime;
this._chunkCount++;
// Get formatted text from the merger (handles SentencePiece spaces correctly)
const texts = this._merger!.getText(this._tokenizer);
const confirmedText = texts.confirmed;
const pendingText = texts.pending;
const fullText = texts.full;
// Notify callbacks
this._callbacks.onConfirmedUpdate?.(confirmedText, mergeResult.confirmed);
this._callbacks.onPendingUpdate?.(pendingText, mergeResult.pending);
this._callbacks.onMergeInfo?.({
lcsLength: mergeResult.lcsLength,
anchorValid: mergeResult.anchorValid,
anchorTokens: mergeResult.anchorTokens
});
// Always log preprocessing info from model metrics
const m = result.metrics;
if (m) {
console.log(`[TokenStreamTranscriber] Chunk #${this._chunkCount + 1}: preprocessor=${m.preprocessor_backend || 'unknown'}, preprocess=${m.preprocess_ms}ms, encode=${m.encode_ms}ms, decode=${m.decode_ms}ms, total=${m.total_ms}ms${m.mel_cache ? `, mel_cache: ${m.mel_cache.cached_frames} cached / ${m.mel_cache.new_frames} new` : ''}`);
}
if (this._config.debug) {
console.log(`[TokenStreamTranscriber] Chunk ${this._chunkCount}: start=${chunkStartTime.toFixed(2)}s, overlap=${actualOverlap.toFixed(2)}s, LCS=${mergeResult.lcsLength}, anchor=${mergeResult.anchorValid}`);
}
return {
confirmedText,
pendingText,
fullText,
lcsLength: mergeResult.lcsLength,
anchorValid: mergeResult.anchorValid,
chunkCount: this._chunkCount,
anchorTokens: mergeResult.anchorTokens,
};
}
/**
* Process a chunk using pre-computed mel features from the mel worker.
* Bypasses the preprocessor entirely — encoder + decoder only.
*
* @param features - Normalized mel features [melBins, T] from mel worker
* @param T - Number of time frames
* @param melBins - Number of mel frequency bins
* @param startTime - Start time of this audio window in seconds
* @param overlapSeconds - Overlap with previous window in seconds (for decoder cache)
*/
async processChunkWithFeatures(
features: Float32Array,
T: number,
melBins: number,
startTime?: number,
overlapSeconds?: number,
): Promise<TokenStreamResult> {
if (!this._merger) {
throw new Error('TokenStreamTranscriber not initialized. Call initialize() first.');
}
if (this._isProcessing) {
if (this._config.debug) {
console.log('[TokenStreamTranscriber] Queued features chunk (processing busy)');
}
return this._getEmptyResult();
}
this._isProcessing = true;
try {
const result = await this._processChunkWithFeaturesInternal(
features, T, melBins, startTime, overlapSeconds,
);
return result;
} catch (error) {
console.error('[TokenStreamTranscriber] Process features chunk error:', error);
this._callbacks.onError?.(error as Error);
throw error;
} finally {
this._isProcessing = false;
}
}
private async _processChunkWithFeaturesInternal(
features: Float32Array,
T: number,
melBins: number,
startTime?: number,
overlapSeconds?: number,
): Promise<TokenStreamResult> {
const model = this._modelManager.getModel();
if (!model) {
throw new Error('Model not available');
}
const chunkStartTime = startTime !== undefined ? startTime : this._currentTimestamp;
// Calculate overlap for decoder cache
let actualOverlap = 0;
if (this._chunkCount > 0 && overlapSeconds !== undefined) {
actualOverlap = overlapSeconds;
} else if (this._chunkCount > 0) {
// Estimate from config
actualOverlap = this._config.overlapDuration;
}
const overlapSec = this._chunkCount > 0 ? actualOverlap : 0;
// Call transcribe with pre-computed features (bypasses preprocessor)
const result = await model.transcribe(null, this._config.sampleRate, {
returnTimestamps: true,
returnTokenIds: true,
returnFrameIndices: true,
returnLogProbs: this._config.returnLogProbs,
timeOffset: chunkStartTime,
frameStride: this._config.frameStride,
precomputedFeatures: { features, T, melBins },
incremental: overlapSec > 0 ? {
cacheKey: 'streaming',
prefixSeconds: overlapSec,
} : undefined,
});
// Merge using LCSPTFAMerger (same as audio path)
const mergeResult = this._merger!.processChunk(
result,
chunkStartTime,
this._chunkCount > 0 ? actualOverlap : 0
);
// Update state for next chunk
this._currentTimestamp = chunkStartTime;
this._chunkCount++;
// Get formatted text
const texts = this._merger!.getText(this._tokenizer);
const confirmedText = texts.confirmed;
const pendingText = texts.pending;
const fullText = texts.full;
// Notify callbacks
this._callbacks.onConfirmedUpdate?.(confirmedText, mergeResult.confirmed);
this._callbacks.onPendingUpdate?.(pendingText, mergeResult.pending);
this._callbacks.onMergeInfo?.({
lcsLength: mergeResult.lcsLength,
anchorValid: mergeResult.anchorValid,
anchorTokens: mergeResult.anchorTokens
});
// Always log preprocessing info from model metrics
const m = result.metrics;
if (m) {
console.log(`[TokenStreamTranscriber] Features chunk #${this._chunkCount + 1}: preprocessor=${m.preprocessor_backend || 'unknown'}, preprocess=${m.preprocess_ms}ms, encode=${m.encode_ms}ms, decode=${m.decode_ms}ms, total=${m.total_ms}ms, T=${T}`);
}
if (this._config.debug) {
console.log(`[TokenStreamTranscriber] Features chunk ${this._chunkCount}: start=${chunkStartTime.toFixed(2)}s, overlap=${actualOverlap.toFixed(2)}s, T=${T}, LCS=${mergeResult.lcsLength}`);
}
return {
confirmedText,
pendingText,
fullText,
lcsLength: mergeResult.lcsLength,
anchorValid: mergeResult.anchorValid,
chunkCount: this._chunkCount,
anchorTokens: mergeResult.anchorTokens,
};
}
/**
* Finalize and get complete transcript.
* Promotes all pending tokens to confirmed.
*/
finalize(): TokenStreamResult {
if (!this._merger) {
return this._getEmptyResult();
}
const texts = this._merger.getText(this._tokenizer);
return {
confirmedText: texts.full, // All text is now confirmed
pendingText: '',
fullText: texts.full,
lcsLength: 0,
anchorValid: true,
chunkCount: this._chunkCount,
};
}
/**
* Reset for a new transcription session.
*/
reset(): void {
if (this._merger) {
this._merger.reset();
}
// Reset mel cache in the model's JS preprocessor (if available)
const model = this._modelManager.getModel();
if (model?.resetMelCache) {
model.resetMelCache();
}
this._currentTimestamp = 0;
this._chunkCount = 0;
this._isProcessing = false;
this._pendingChunk = null;
}
/**
* Get current state for debugging.
*/
getState(): any {
return {
chunkCount: this._chunkCount,
currentTimestamp: this._currentTimestamp,
isProcessing: this._isProcessing,
mergerState: this._merger?.getState() || null,
};
}
/**
* Check if currently processing.
*/
isProcessing(): boolean {
return this._isProcessing;
}
private _getEmptyResult(): TokenStreamResult {
return {
confirmedText: '',
pendingText: '',
fullText: '',
lcsLength: 0,
anchorValid: false,
chunkCount: this._chunkCount,
};
}
/**
* Connect to AudioEngine for automatic window-based streaming.
*
* Usage:
* ```
* const unsubscribe = transcriber.connectToAudioEngine(audioEngine);
* // ... later
* unsubscribe();
* ```
*
* @param audioEngine - AudioEngine instance with onWindowChunk support
* @returns Unsubscribe function
*/
connectToAudioEngine(audioEngine: any): () => void {
if (!audioEngine.onWindowChunk) {
throw new Error('AudioEngine does not support onWindowChunk. Update to v3.0.');
}
return audioEngine.onWindowChunk(
this._config.windowDuration,
this._config.overlapDuration,
this._config.windowDuration - this._config.overlapDuration, // Default trigger interval
async (audio: Float32Array, startTime: number) => {
try {
await this.processChunk(audio, startTime);
} catch (e) {
console.error('[TokenStreamTranscriber] Window processing error:', e);
}
}
);
}
/**
* Get configuration.
*/
getConfig(): Required<TokenStreamConfig> {
return { ...this._config };
}
}
|