File size: 19,610 Bytes
c69c1ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
// Svara TTS WebGPU worker.
//
// Architecture:
//   1) Llama-3.2-3B causal LM (loaded via @huggingface/transformers v4) emits
//      audio token IDs in the range [128266, 156938).
//   2) We group every 7-token bundle into a SNAC frame.
//   3) Offline decode mirrors Kenpath's streaming path: decode a sliding
//      4-frame SNAC window and keep samples [2048:4096] from each window.
//      That matches the codec's stable synthesis region and avoids the
//      "behind a fan" smear seen when decoding the whole sequence in one shot.

import {
  AutoTokenizer,
  AutoModelForCausalLM,
  LogitsProcessor,
  LogitsProcessorList,
  Tensor,
} from "@huggingface/transformers";
import * as ort from "onnxruntime-web/webgpu";

// ORT-Web's .wasm/.mjs files aren't served by Vite by default; vite.config.js
// copies them from node_modules to /ort-wasm/ via vite-plugin-static-copy.
ort.env.wasm.wasmPaths = "/ort-wasm/";

// --- WebGPU feature detection -----------------------------------------------
let fp16_supported = false;
try {
  const adapter = await navigator.gpu.requestAdapter();
  if (!adapter) throw new Error("WebGPU is not supported (no adapter found)");
  fp16_supported = adapter.features.has("shader-f16");
  self.postMessage({ status: "feature-success", fp16: fp16_supported });
} catch (e) {
  self.postMessage({ status: "feature-error", data: e.toString() });
  throw e;
}

// --- Constants matching upstream Svara inference -----------------------------
const EOT = 128009;
const SOS = 128257, EOS = 128258;
const SOH = 128259, EOH = 128260;
const SOAI = 128261;
const AUDIO_OFFSET = 128266;
const AUDIO_END = AUDIO_OFFSET + 7 * 4096;
const WINDOW_FRAMES = 4;
const WINDOW_AUDIO_START = 2048;
const WINDOW_AUDIO_END = 4096;
const SAMPLE_RATE = 24000;

const SVARA_REPO = "shreyask/svara-tts-v1-ONNX";
const SNAC_REPO = "onnx-community/snac_24khz-ONNX";
const SUPPORTED_DTYPES = new Set(["q4f16", "q8"]);

// Lazy load the tokenizer once -- it's the same across dtypes.
let tokenizerPromise = null;
function getTokenizer() {
  return (tokenizerPromise ??= AutoTokenizer.from_pretrained(SVARA_REPO));
}

// SNAC decoder is small (~26 MB at fp16); load once, share across LM dtypes.
let snacPromise = null;
function getSnac() {
  return (snacPromise ??= (async () => {
    const url = `https://huggingface.co/${SNAC_REPO}/resolve/main/onnx/decoder_model${fp16_supported ? "_fp16" : ""}.onnx`;
    return ort.InferenceSession.create(url, { executionProviders: ["webgpu"] });
  })());
}

// LM is per-dtype. Cache by dtype string so switching back is instant.
const lmCache = new Map();
function getLM(dtype) {
  if (!lmCache.has(dtype)) {
    lmCache.set(
      dtype,
      AutoModelForCausalLM.from_pretrained(SVARA_REPO, {
        dtype,
        device: "webgpu",
        // Number of external data chunks to fetch alongside the .onnx graph.
        // q4f16 is one .onnx_data file; q8 is sharded into 3 chunks
        // (.onnx_data, _data_1, _data_2) to stay under the ~2 GB browser
        // ArrayBuffer ceiling. transformers.js v4 accepts a number here per
        // its types: `false` | `true` (=1) | <number of chunks>.
        use_external_data_format: dtype === "q8" ? 3 : true,
      }),
    );
  }
  return lmCache.get(dtype);
}

// --- Generation guards ------------------------------------------------------
// Svara should only emit 7-band audio tokens followed by END_OF_SPEECH. If we
// let the sampler wander into the text/control vocab, the rest of the clip
// turns phasey/robotic because frame alignment is lost.
class SvaraLogitsProcessor extends LogitsProcessor {
  constructor(promptLength) {
    super();
    this.promptLength = promptLength;
  }

  _call(inputIds, logits) {
    for (let i = 0; i < inputIds.length; i++) {
      const data = logits[i].data;
      const step = inputIds[i].length - this.promptLength;

      if (step === 0) {
        data.fill(-Infinity);
        data[SOAI] = 0;
        continue;
      }

      if (step === 1) {
        data.fill(-Infinity);
        data[SOS] = 0;
        continue;
      }

      const eosLogit = data[EOS];
      data.subarray(0, AUDIO_OFFSET).fill(-Infinity);
      data.subarray(AUDIO_END).fill(-Infinity);
      data[EOS] = eosLogit;
    }
    return logits;
  }
}

function buildLogitsProcessor(promptLength) {
  const list = new LogitsProcessorList();
  list.push(new SvaraLogitsProcessor(promptLength));
  return list;
}

function estimateAudioTokenBudget(text) {
  const spokenText = stripTrailingEmotionTag(text);
  const graphemeCount = Array.from(
    new Intl.Segmenter(undefined, { granularity: "grapheme" }).segment(spokenText),
    ({ segment }) => segment,
  ).filter((segment) => /\S/u.test(segment)).length;
  const punctuationGroups = Array.from(
    spokenText.matchAll(/[.,!?;:।॥…\-—]+/gu),
  ).length;
  const wordCount = spokenText.split(/\s+/u).filter(Boolean).length;

  const roughBudget = graphemeCount * 12 + wordCount * 20 + punctuationGroups * 28 + 84;
  const clampedBudget = Math.max(224, Math.min(1120, roughBudget));
  return Math.ceil(clampedBudget / 7) * 7;
}

function getTrailingEmotionTag(text) {
  return text.match(/\s*(<[^>]+>)\s*$/u)?.[1] ?? "";
}

function stripTrailingEmotionTag(text) {
  return text.replace(/\s*<[^>]+>\s*$/u, "").trim();
}

function normalizeTextForSvara(text) {
  return text
    .replace(/\.{2,}/gu, ",")
    .replace(/…+/gu, ",")
    .replace(/[—–]+/gu, ",")
    .replace(/\s+/gu, " ")
    .replace(/\s*([,.;!?।॥])\s*/gu, "$1 ")
    .trim();
}

function countChunkGraphemes(chunk) {
  return Array.from(
    new Intl.Segmenter(undefined, { granularity: "grapheme" }).segment(chunk),
    ({ segment }) => segment,
  ).filter((segment) => /\S/u.test(segment)).length;
}

function countChunkWords(chunk) {
  return chunk.split(/\s+/u).filter(Boolean).length;
}

function splitLongChunk(chunk) {
  const graphemeCount = countChunkGraphemes(chunk);
  const wordCount = countChunkWords(chunk);

  if (graphemeCount <= 28 || wordCount <= 5) return [chunk];

  const parts = chunk.split(/\s*,\s*/u).map((part) => part.trim()).filter(Boolean);
  return parts.length > 1 ? parts : [chunk];
}

function mergeTinyChunks(chunks) {
  const merged = [];

  for (const chunk of chunks) {
    const graphemeCount = countChunkGraphemes(chunk);
    const wordCount = countChunkWords(chunk);
    const shouldAttach =
      merged.length > 0 &&
      !/[.!?।॥]$/u.test(merged.at(-1)) &&
      (graphemeCount < 10 || wordCount < 3);

    if (shouldAttach) {
      merged[merged.length - 1] = `${merged.at(-1)}, ${chunk}`;
      continue;
    }

    merged.push(chunk);
  }

  return merged;
}

function splitTextForSvara(text) {
  const emotionTag = getTrailingEmotionTag(text);
  const spokenText = normalizeTextForSvara(stripTrailingEmotionTag(text));

  if (!spokenText) return [];

  const chunks = mergeTinyChunks(
    spokenText
      .match(/[^.!?।॥]+[.!?।॥]?/gu)
      ?.map((part) => part.trim())
      .filter(Boolean)
      .flatMap(splitLongChunk) ?? [],
  );

  if (!emotionTag) return chunks;
  return chunks.map((chunk, index) =>
    index === chunks.length - 1 ? `${chunk} ${emotionTag}` : chunk,
  );
}

function mergeTinyLeadingChunks(chunks) {
  const merged = [];

  for (let i = 0; i < chunks.length; i++) {
    const chunk = chunks[i];
    const graphemeCount = countChunkGraphemes(chunk);
    const wordCount = countChunkWords(chunk);

    if (graphemeCount < 10 && wordCount < 3) {
      if (i + 1 < chunks.length) {
        chunks[i + 1] = `${chunk}, ${chunks[i + 1]}`;
        continue;
      }
      if (merged.length > 0) {
        merged[merged.length - 1] = `${merged.at(-1)}, ${chunk}`;
        continue;
      }
    }

    merged.push(chunk);
  }

  return merged;
}

function splitEmotionSafeTextForSvara(text) {
  const emotionTag = getTrailingEmotionTag(text);
  const spokenText = normalizeTextForSvara(stripTrailingEmotionTag(text));

  if (!spokenText) return [];

  const chunks = spokenText
    .match(/[^.!?।॥]+[.!?।॥]?/gu)
    ?.map((part) => part.trim())
    .filter(Boolean)
    .flatMap((sentence) => {
      const commaParts = sentence
        .split(/\s*,\s*/u)
        .map((part) => part.trim())
        .filter(Boolean);
      return mergeTinyLeadingChunks(commaParts);
    }) ?? [];

  if (!emotionTag) return chunks;
  return chunks.map((chunk, index) =>
    index === chunks.length - 1 ? `${chunk} ${emotionTag}` : chunk,
  );
}

function splitFinalEmotionClauseTextForSvara(text) {
  const emotionTag = getTrailingEmotionTag(text);
  const spokenText = normalizeTextForSvara(stripTrailingEmotionTag(text));

  if (!spokenText) return [];

  const chunks = mergeTinyLeadingChunks(
    spokenText.split(/\s*,\s*/u).map((part) => part.trim()).filter(Boolean),
  );

  if (!emotionTag) return chunks;
  return chunks.map((chunk, index) =>
    index === chunks.length - 1 ? `${chunk} ${emotionTag}` : chunk,
  );
}

function buildPromptVariants(text) {
  const rawText = text.trim();
  const spokenText = normalizeTextForSvara(stripTrailingEmotionTag(text));
  if (!rawText && !spokenText) return [];

  const variants = rawText ? [[rawText]] : [];
  variants.push(
    splitTextForSvara(text),
    splitEmotionSafeTextForSvara(text),
    splitFinalEmotionClauseTextForSvara(text),
  );

  if (getTrailingEmotionTag(text)) {
    variants.push([spokenText]);
    variants.push(splitEmotionSafeTextForSvara(spokenText));
  }

  const seen = new Set();
  return variants.filter((chunks) => {
    if (chunks.length === 0) return false;
    const key = chunks.join("\u241e");
    if (seen.has(key)) return false;
    seen.add(key);
    return true;
  });
}

function pauseDurationForChunk(chunk, isLast) {
  if (isLast) return 0;
  const trimmed = chunk.trim();
  if (/[!?]$/u.test(trimmed)) return 0.26;
  if (/[.]$/u.test(trimmed)) return 0.3;
  return 0.18;
}

function concatFloat32Arrays(chunks) {
  const totalLength = chunks.reduce((sum, chunk) => sum + chunk.length, 0);
  const merged = new Float32Array(totalLength);
  let offset = 0;
  for (const chunk of chunks) {
    merged.set(chunk, offset);
    offset += chunk.length;
  }
  return merged;
}

function pcmStats(samples) {
  let peak = 0;
  let sumSquares = 0;

  for (let i = 0; i < samples.length; i++) {
    const value = Math.abs(samples[i]);
    if (value > peak) peak = value;
    sumSquares += value * value;
  }

  const rms = samples.length > 0 ? Math.sqrt(sumSquares / samples.length) : 0;
  return { peak, rms };
}

function isNearlySilent(samples) {
  const { peak, rms } = pcmStats(samples);
  return peak < 0.006 && rms < 0.0015;
}

function isComplexQ4Prompt(text) {
  const spokenText = stripTrailingEmotionTag(text);
  const wordCount = countChunkWords(spokenText);
  const punctuationGroups = Array.from(
    spokenText.matchAll(/[.,!?;:।॥…\-—]+/gu),
  ).length;
  return punctuationGroups >= 3 || wordCount >= 8 || (
    getTrailingEmotionTag(text) && punctuationGroups >= 1 && wordCount >= 5
  );
}

async function synthesizeChunks(tokenizer, lm, speaker_id, chunks, generation) {
  const pcmChunks = [];

  for (let index = 0; index < chunks.length; index++) {
    const chunk = chunks[index];
    const promptIds = buildPrompt(tokenizer, chunk, speaker_id);
    const inputIds = new Tensor(
      "int64",
      BigInt64Array.from(promptIds.map(BigInt)),
      [1, promptIds.length],
    );

    const maxAudioTokens = estimateAudioTokenBudget(chunk);
    const out = await lm.generate({
      inputs: inputIds,
      max_new_tokens: maxAudioTokens + 3,
      logits_processor: buildLogitsProcessor(promptIds.length),
      ...generation,
      repetition_penalty: 1.0,
      eos_token_id: EOS,
    });

    const allIds = Array.from(out.data, (x) => Number(x));
    const audioIds = extractAudioTokens(allIds, promptIds.length);
    if (audioIds.length === 0) {
      throw new Error(`LM produced no audio tokens for chunk ${index + 1}/${chunks.length}.`);
    }

    const pcm = await decodeSnacStable(audioIds);
    pcmChunks.push(pcm);

    const pauseSeconds = pauseDurationForChunk(chunk, index === chunks.length - 1);
    if (pauseSeconds > 0) {
      pcmChunks.push(new Float32Array(Math.round(SAMPLE_RATE * pauseSeconds)));
    }
  }

  return concatFloat32Arrays(pcmChunks);
}

// --- Token-stream → SNAC code conversion ------------------------------------
// Reference: mlx_audio/tts/models/llama/llama.py:codes_to_layers
//   layer_1 (band 0):           [c0]                — 1 code per coarse frame
//   layer_2 (bands 1, 4):       [c1, c4]            — 2 codes per coarse frame
//   layer_3 (bands 2, 3, 5, 6): [c2, c3, c5, c6]    — 4 codes per coarse frame
function codesToLayers(audioTokenIds) {
  const N = Math.floor(audioTokenIds.length / 7);
  const l1 = new BigInt64Array(N);
  const l2 = new BigInt64Array(N * 2);
  const l3 = new BigInt64Array(N * 4);
  for (let i = 0; i < N; i++) {
    const base = i * 7;
    l1[i]         = BigInt(audioTokenIds[base    ] - AUDIO_OFFSET - 0 * 4096);
    l2[2 * i + 0] = BigInt(audioTokenIds[base + 1] - AUDIO_OFFSET - 1 * 4096);
    l3[4 * i + 0] = BigInt(audioTokenIds[base + 2] - AUDIO_OFFSET - 2 * 4096);
    l3[4 * i + 1] = BigInt(audioTokenIds[base + 3] - AUDIO_OFFSET - 3 * 4096);
    l2[2 * i + 1] = BigInt(audioTokenIds[base + 4] - AUDIO_OFFSET - 4 * 4096);
    l3[4 * i + 2] = BigInt(audioTokenIds[base + 5] - AUDIO_OFFSET - 5 * 4096);
    l3[4 * i + 3] = BigInt(audioTokenIds[base + 6] - AUDIO_OFFSET - 6 * 4096);
  }
  return { l1, l2, l3, N };
}

async function decodeSnacWindow(audioTokenIds) {
  const snac = await getSnac();
  const { l1, l2, l3, N } = codesToLayers(audioTokenIds);
  const feeds = {
    [snac.inputNames[0]]: new ort.Tensor("int64", l1, [1, N]),
    [snac.inputNames[1]]: new ort.Tensor("int64", l2, [1, N * 2]),
    [snac.inputNames[2]]: new ort.Tensor("int64", l3, [1, N * 4]),
  };
  const out = await snac.run(feeds);
  return out[snac.outputNames[0]].data;
}

async function decodeSnacStable(audioTokenIds) {
  const numFrames = Math.floor(audioTokenIds.length / 7);
  if (numFrames === 0) return new Float32Array(0);

  if (numFrames < WINDOW_FRAMES) {
    return await decodeSnacWindow(audioTokenIds);
  }

  const chunks = [];
  let totalLength = 0;

  for (let start = 0; start <= numFrames - WINDOW_FRAMES; start++) {
    const windowIds = audioTokenIds.slice(start * 7, (start + WINDOW_FRAMES) * 7);
    const decoded = await decodeSnacWindow(windowIds);
    const stable = decoded.slice(WINDOW_AUDIO_START, WINDOW_AUDIO_END);
    chunks.push(stable);
    totalLength += stable.length;
  }

  const merged = new Float32Array(totalLength);
  let offset = 0;
  for (const chunk of chunks) {
    merged.set(chunk, offset);
    offset += chunk.length;
  }
  return merged;
}

// Match the exported ONNX repo README:
//   [SOH, BOS, "<voice>: <text>" tokens, EOT, EOH]
// The model predicts SOAI -> SOS -> audio tokens -> EOS itself.
function buildPrompt(tokenizer, text, voice) {
  const body = tokenizer.encode(`${voice}: ${text}`, { add_special_tokens: false });
  return [SOH, tokenizer.bos_token_id, ...body, EOT, EOH];
}

// Keep audio tokens after the first START_OF_SPEECH emitted by the model.
function extractAudioTokens(allTokenIds, promptLength) {
  let sosIdx = -1;
  for (let i = promptLength; i < allTokenIds.length; i++) {
    if (allTokenIds[i] === SOS) {
      sosIdx = i;
      break;
    }
  }
  if (sosIdx === -1) return [];

  const audio = [];
  for (let i = sosIdx + 1; i < allTokenIds.length; i++) {
    const tokenId = allTokenIds[i];
    if (tokenId === EOS) break;
    if (tokenId >= AUDIO_OFFSET && tokenId < AUDIO_END) {
      audio.push(tokenId);
    }
  }
  return audio.slice(0, audio.length - (audio.length % 7));
}

// --- WAV encoder (24 kHz, mono, PCM16) --------------------------------------
function pcmFloat32ToWav(samples, sampleRate) {
  const bufLen = 44 + samples.length * 2;
  const buf = new ArrayBuffer(bufLen);
  const v = new DataView(buf);
  let p = 0;
  const w = (s) => { for (let i = 0; i < s.length; i++) v.setUint8(p++, s.charCodeAt(i)); };
  w("RIFF");
  v.setUint32(p, 36 + samples.length * 2, true); p += 4;
  w("WAVEfmt ");
  v.setUint32(p, 16, true); p += 4;
  v.setUint16(p, 1, true); p += 2;
  v.setUint16(p, 1, true); p += 2;
  v.setUint32(p, sampleRate, true); p += 4;
  v.setUint32(p, sampleRate * 2, true); p += 4;
  v.setUint16(p, 2, true); p += 2;
  v.setUint16(p, 16, true); p += 2;
  w("data");
  v.setUint32(p, samples.length * 2, true); p += 4;
  for (let i = 0; i < samples.length; i++) {
    const s = Math.max(-1, Math.min(1, samples[i]));
    v.setInt16(p, s < 0 ? s * 0x8000 : s * 0x7fff, true);
    p += 2;
  }
  return buf;
}

// --- Sampling defaults per dtype --------------------------------------------
// Transformers.js v4 currently ignores top-k/top-p on this path, so unconstrained
// sampling drifts badly on quantized Svara and turns later words robotic. Use
// greedy decoding by default for stability; q8 can tolerate a little sampling.
function generationFor(dtype) {
  return dtype === "q8"
    ? { do_sample: true, temperature: 0.35, min_new_tokens: 30 }
    : { do_sample: false, min_new_tokens: 30 };
}

function generationPlansFor(dtype, text) {
  const base = generationFor(dtype);
  if (dtype !== "q4f16" || !isComplexQ4Prompt(text)) {
    return [base];
  }

  return [
    {
      do_sample: true,
      temperature: 0.6,
      top_k: 40,
      top_p: 0.9,
      min_new_tokens: 30,
    },
    base,
  ];
}

// --- Message handler --------------------------------------------------------
self.addEventListener("message", async (e) => {
  const { type, text, speaker_id, dtype: requested } = e.data;
  const dtype = SUPPORTED_DTYPES.has(requested) ? requested : "q4f16";

  try {
    if (type === "preload") {
      // Triggered by the explicit "Load model" action in the UI.
      self.postMessage({ status: "loading", dtype });
      await Promise.all([getTokenizer(), getSnac(), getLM(dtype)]);
      self.postMessage({ status: "ready", dtype });
      return;
    }

    self.postMessage({ status: "loading", dtype });
    const [tokenizer, lm] = await Promise.all([getTokenizer(), getLM(dtype)]);
    await getSnac(); // warm

    const variants = buildPromptVariants(text);
    if (variants.length === 0) {
      throw new Error("No speakable text found after normalization.");
    }

    const generations = generationPlansFor(dtype, text);
    let mergedPcm = null;
    let lastError = null;

    for (const generation of generations) {
      for (const chunks of variants) {
        try {
          const candidate = await synthesizeChunks(
            tokenizer,
            lm,
            speaker_id,
            chunks,
            generation,
          );
          if (isNearlySilent(candidate)) {
            lastError = new Error("Generated near-silent audio.");
            continue;
          }
          mergedPcm = candidate;
          break;
        } catch (err) {
          lastError = err;
        }
      }
      if (mergedPcm) {
        break;
      }
    }

    if (!mergedPcm) {
      throw lastError ?? new Error("Synthesis failed for all prompt variants.");
    }

    const wav = pcmFloat32ToWav(mergedPcm, SAMPLE_RATE);
    const blob = new Blob([wav], { type: "audio/wav" });
    self.postMessage({
      status: "complete",
      audio: URL.createObjectURL(blob),
      text,
      voice: speaker_id,
      dtype,
    });
  } catch (err) {
    self.postMessage({ status: "error", data: String(err), dtype });
    console.error(err);
  }
});