562 lines
19 KiB
JavaScript
562 lines
19 KiB
JavaScript
import * as ort from 'onnxruntime-web';
|
||
|
||
// Available languages for multilingual TTS
|
||
export const AVAILABLE_LANGS = ['en', 'ko', 'es', 'pt', 'fr'];
|
||
|
||
export function isValidLang(lang) {
|
||
return AVAILABLE_LANGS.includes(lang);
|
||
}
|
||
|
||
/**
|
||
* Unicode Text Processor
|
||
*/
|
||
export class UnicodeProcessor {
|
||
constructor(indexer) {
|
||
this.indexer = indexer;
|
||
}
|
||
|
||
call(textList, langList) {
|
||
const processedTexts = textList.map((text, i) => this.preprocessText(text, langList[i]));
|
||
|
||
const textIdsLengths = processedTexts.map(text => text.length);
|
||
const maxLen = Math.max(...textIdsLengths);
|
||
|
||
const textIds = processedTexts.map(text => {
|
||
const row = new Array(maxLen).fill(0);
|
||
for (let j = 0; j < text.length; j++) {
|
||
const codePoint = text.codePointAt(j);
|
||
row[j] = (codePoint < this.indexer.length) ? this.indexer[codePoint] : -1;
|
||
}
|
||
return row;
|
||
});
|
||
|
||
const textMask = this.getTextMask(textIdsLengths);
|
||
return { textIds, textMask };
|
||
}
|
||
|
||
preprocessText(text, lang) {
|
||
// TODO: Need advanced normalizer for better performance
|
||
text = text.normalize('NFKD');
|
||
|
||
// Remove emojis (wide Unicode range)
|
||
const emojiPattern = /[\u{1F600}-\u{1F64F}\u{1F300}-\u{1F5FF}\u{1F680}-\u{1F6FF}\u{1F700}-\u{1F77F}\u{1F780}-\u{1F7FF}\u{1F800}-\u{1F8FF}\u{1F900}-\u{1F9FF}\u{1FA00}-\u{1FA6F}\u{1FA70}-\u{1FAFF}\u{2600}-\u{26FF}\u{2700}-\u{27BF}\u{1F1E6}-\u{1F1FF}]+/gu;
|
||
text = text.replace(emojiPattern, '');
|
||
|
||
// Replace various dashes and symbols
|
||
const replacements = {
|
||
'–': '-',
|
||
'‑': '-',
|
||
'—': '-',
|
||
'_': ' ',
|
||
'\u201C': '"', // left double quote "
|
||
'\u201D': '"', // right double quote "
|
||
'\u2018': "'", // left single quote '
|
||
'\u2019': "'", // right single quote '
|
||
'´': "'",
|
||
'`': "'",
|
||
'[': ' ',
|
||
']': ' ',
|
||
'|': ' ',
|
||
'/': ' ',
|
||
'#': ' ',
|
||
'→': ' ',
|
||
'←': ' ',
|
||
};
|
||
for (const [k, v] of Object.entries(replacements)) {
|
||
text = text.replaceAll(k, v);
|
||
}
|
||
|
||
// Remove special symbols
|
||
text = text.replace(/[♥☆♡©\\]/g, '');
|
||
|
||
// Replace known expressions
|
||
const exprReplacements = {
|
||
'@': ' at ',
|
||
'e.g.,': 'for example, ',
|
||
'i.e.,': 'that is, ',
|
||
};
|
||
for (const [k, v] of Object.entries(exprReplacements)) {
|
||
text = text.replaceAll(k, v);
|
||
}
|
||
|
||
// Fix spacing around punctuation
|
||
text = text.replace(/ ,/g, ',');
|
||
text = text.replace(/ \./g, '.');
|
||
text = text.replace(/ !/g, '!');
|
||
text = text.replace(/ \?/g, '?');
|
||
text = text.replace(/ ;/g, ';');
|
||
text = text.replace(/ :/g, ':');
|
||
text = text.replace(/ '/g, "'");
|
||
|
||
// Remove duplicate quotes
|
||
while (text.includes('""')) {
|
||
text = text.replace('""', '"');
|
||
}
|
||
while (text.includes("''")) {
|
||
text = text.replace("''", "'");
|
||
}
|
||
while (text.includes('``')) {
|
||
text = text.replace('``', '`');
|
||
}
|
||
|
||
// Remove extra spaces
|
||
text = text.replace(/\s+/g, ' ').trim();
|
||
|
||
// If text doesn't end with punctuation, quotes, or closing brackets, add a period
|
||
if (!/[.!?;:,'\"')\]}…。」』】〉》›»]$/.test(text)) {
|
||
text += '.';
|
||
}
|
||
|
||
// Validate language
|
||
if (!isValidLang(lang)) {
|
||
throw new Error(`Invalid language: ${lang}. Available: ${AVAILABLE_LANGS.join(', ')}`);
|
||
}
|
||
|
||
// Wrap text with language tags
|
||
text = `<${lang}>${text}</${lang}>`;
|
||
|
||
return text;
|
||
}
|
||
|
||
getTextMask(textIdsLengths) {
|
||
const maxLen = Math.max(...textIdsLengths);
|
||
return this.lengthToMask(textIdsLengths, maxLen);
|
||
}
|
||
|
||
lengthToMask(lengths, maxLen = null) {
|
||
const actualMaxLen = maxLen || Math.max(...lengths);
|
||
return lengths.map(len => {
|
||
const row = new Array(actualMaxLen).fill(0.0);
|
||
for (let j = 0; j < Math.min(len, actualMaxLen); j++) {
|
||
row[j] = 1.0;
|
||
}
|
||
return [row];
|
||
});
|
||
}
|
||
}
|
||
|
||
/**
|
||
* Style class to hold TTL and DP tensors
|
||
*/
|
||
export class Style {
|
||
constructor(ttlTensor, dpTensor) {
|
||
this.ttl = ttlTensor;
|
||
this.dp = dpTensor;
|
||
}
|
||
}
|
||
|
||
/**
|
||
* Text-to-Speech class
|
||
*/
|
||
export class TextToSpeech {
|
||
constructor(cfgs, textProcessor, dpOrt, textEncOrt, vectorEstOrt, vocoderOrt) {
|
||
this.cfgs = cfgs;
|
||
this.textProcessor = textProcessor;
|
||
this.dpOrt = dpOrt;
|
||
this.textEncOrt = textEncOrt;
|
||
this.vectorEstOrt = vectorEstOrt;
|
||
this.vocoderOrt = vocoderOrt;
|
||
this.sampleRate = cfgs.ae.sample_rate;
|
||
}
|
||
|
||
async _infer(textList, langList, style, totalStep, speed = 1.05, progressCallback = null) {
|
||
const bsz = textList.length;
|
||
|
||
// Process text
|
||
const { textIds, textMask } = this.textProcessor.call(textList, langList);
|
||
|
||
const textIdsFlat = new BigInt64Array(textIds.flat().map(x => BigInt(x)));
|
||
const textIdsShape = [bsz, textIds[0].length];
|
||
const textIdsTensor = new ort.Tensor('int64', textIdsFlat, textIdsShape);
|
||
|
||
const textMaskFlat = new Float32Array(textMask.flat(2));
|
||
const textMaskShape = [bsz, 1, textMask[0][0].length];
|
||
const textMaskTensor = new ort.Tensor('float32', textMaskFlat, textMaskShape);
|
||
|
||
// Predict duration
|
||
const dpOutputs = await this.dpOrt.run({
|
||
text_ids: textIdsTensor,
|
||
style_dp: style.dp,
|
||
text_mask: textMaskTensor
|
||
});
|
||
const duration = Array.from(dpOutputs.duration.data);
|
||
|
||
// Apply speed factor to duration
|
||
for (let i = 0; i < duration.length; i++) {
|
||
duration[i] /= speed;
|
||
}
|
||
|
||
// Encode text
|
||
const textEncOutputs = await this.textEncOrt.run({
|
||
text_ids: textIdsTensor,
|
||
style_ttl: style.ttl,
|
||
text_mask: textMaskTensor
|
||
});
|
||
const textEmb = textEncOutputs.text_emb;
|
||
|
||
// Sample noisy latent
|
||
let { xt, latentMask } = this.sampleNoisyLatent(
|
||
duration,
|
||
this.sampleRate,
|
||
this.cfgs.ae.base_chunk_size,
|
||
this.cfgs.ttl.chunk_compress_factor,
|
||
this.cfgs.ttl.latent_dim
|
||
);
|
||
|
||
const latentMaskFlat = new Float32Array(latentMask.flat(2));
|
||
const latentMaskShape = [bsz, 1, latentMask[0][0].length];
|
||
const latentMaskTensor = new ort.Tensor('float32', latentMaskFlat, latentMaskShape);
|
||
|
||
// Prepare constant arrays
|
||
const totalStepArray = new Float32Array(bsz).fill(totalStep);
|
||
const totalStepTensor = new ort.Tensor('float32', totalStepArray, [bsz]);
|
||
|
||
// Denoising loop
|
||
for (let step = 0; step < totalStep; step++) {
|
||
if (progressCallback) {
|
||
progressCallback(step + 1, totalStep);
|
||
}
|
||
|
||
const currentStepArray = new Float32Array(bsz).fill(step);
|
||
const currentStepTensor = new ort.Tensor('float32', currentStepArray, [bsz]);
|
||
|
||
const xtFlat = new Float32Array(xt.flat(2));
|
||
const xtShape = [bsz, xt[0].length, xt[0][0].length];
|
||
const xtTensor = new ort.Tensor('float32', xtFlat, xtShape);
|
||
|
||
const vectorEstOutputs = await this.vectorEstOrt.run({
|
||
noisy_latent: xtTensor,
|
||
text_emb: textEmb,
|
||
style_ttl: style.ttl,
|
||
latent_mask: latentMaskTensor,
|
||
text_mask: textMaskTensor,
|
||
current_step: currentStepTensor,
|
||
total_step: totalStepTensor
|
||
});
|
||
|
||
const denoised = Array.from(vectorEstOutputs.denoised_latent.data);
|
||
|
||
// Reshape to 3D
|
||
const latentDim = xt[0].length;
|
||
const latentLen = xt[0][0].length;
|
||
xt = [];
|
||
let idx = 0;
|
||
for (let b = 0; b < bsz; b++) {
|
||
const batch = [];
|
||
for (let d = 0; d < latentDim; d++) {
|
||
const row = [];
|
||
for (let t = 0; t < latentLen; t++) {
|
||
row.push(denoised[idx++]);
|
||
}
|
||
batch.push(row);
|
||
}
|
||
xt.push(batch);
|
||
}
|
||
}
|
||
|
||
// Generate waveform
|
||
const finalXtFlat = new Float32Array(xt.flat(2));
|
||
const finalXtShape = [bsz, xt[0].length, xt[0][0].length];
|
||
const finalXtTensor = new ort.Tensor('float32', finalXtFlat, finalXtShape);
|
||
|
||
const vocoderOutputs = await this.vocoderOrt.run({
|
||
latent: finalXtTensor
|
||
});
|
||
|
||
const wav = Array.from(vocoderOutputs.wav_tts.data);
|
||
|
||
return { wav, duration };
|
||
}
|
||
|
||
async call(text, lang, style, totalStep, speed = 1.05, silenceDuration = 0.3, progressCallback = null) {
|
||
if (style.ttl.dims[0] !== 1) {
|
||
throw new Error('Single speaker text to speech only supports single style');
|
||
}
|
||
const maxLen = lang === 'ko' ? 120 : 300;
|
||
const textList = chunkText(text, maxLen);
|
||
const langList = new Array(textList.length).fill(lang);
|
||
let wavCat = [];
|
||
let durCat = 0;
|
||
|
||
for (let i = 0; i < textList.length; i++) {
|
||
const { wav, duration } = await this._infer([textList[i]], [langList[i]], style, totalStep, speed, progressCallback);
|
||
|
||
if (wavCat.length === 0) {
|
||
wavCat = wav;
|
||
durCat = duration[0];
|
||
} else {
|
||
const silenceLen = Math.floor(silenceDuration * this.sampleRate);
|
||
const silence = new Array(silenceLen).fill(0);
|
||
wavCat = [...wavCat, ...silence, ...wav];
|
||
durCat += duration[0] + silenceDuration;
|
||
}
|
||
}
|
||
|
||
return { wav: wavCat, duration: [durCat] };
|
||
}
|
||
|
||
async batch(textList, langList, style, totalStep, speed = 1.05, progressCallback = null) {
|
||
return await this._infer(textList, langList, style, totalStep, speed, progressCallback);
|
||
}
|
||
|
||
sampleNoisyLatent(duration, sampleRate, baseChunkSize, chunkCompress, latentDim) {
|
||
const bsz = duration.length;
|
||
const maxDur = Math.max(...duration);
|
||
|
||
const wavLenMax = Math.floor(maxDur * sampleRate);
|
||
const wavLengths = duration.map(d => Math.floor(d * sampleRate));
|
||
|
||
const chunkSize = baseChunkSize * chunkCompress;
|
||
const latentLen = Math.floor((wavLenMax + chunkSize - 1) / chunkSize);
|
||
const latentDimVal = latentDim * chunkCompress;
|
||
|
||
const xt = [];
|
||
for (let b = 0; b < bsz; b++) {
|
||
const batch = [];
|
||
for (let d = 0; d < latentDimVal; d++) {
|
||
const row = [];
|
||
for (let t = 0; t < latentLen; t++) {
|
||
// Box-Muller transform
|
||
const u1 = Math.max(0.0001, Math.random());
|
||
const u2 = Math.random();
|
||
const val = Math.sqrt(-2.0 * Math.log(u1)) * Math.cos(2.0 * Math.PI * u2);
|
||
row.push(val);
|
||
}
|
||
batch.push(row);
|
||
}
|
||
xt.push(batch);
|
||
}
|
||
|
||
const latentLengths = wavLengths.map(len => Math.floor((len + chunkSize - 1) / chunkSize));
|
||
const latentMask = this.lengthToMask(latentLengths, latentLen);
|
||
|
||
// Apply mask
|
||
for (let b = 0; b < bsz; b++) {
|
||
for (let d = 0; d < latentDimVal; d++) {
|
||
for (let t = 0; t < latentLen; t++) {
|
||
xt[b][d][t] *= latentMask[b][0][t];
|
||
}
|
||
}
|
||
}
|
||
|
||
return { xt, latentMask };
|
||
}
|
||
|
||
lengthToMask(lengths, maxLen = null) {
|
||
const actualMaxLen = maxLen || Math.max(...lengths);
|
||
return lengths.map(len => {
|
||
const row = new Array(actualMaxLen).fill(0.0);
|
||
for (let j = 0; j < Math.min(len, actualMaxLen); j++) {
|
||
row[j] = 1.0;
|
||
}
|
||
return [row];
|
||
});
|
||
}
|
||
}
|
||
|
||
/**
|
||
* Load voice style from JSON files
|
||
*/
|
||
export async function loadVoiceStyle(voiceStylePaths, verbose = false) {
|
||
const bsz = voiceStylePaths.length;
|
||
|
||
// Read first file to get dimensions
|
||
const firstResponse = await fetch(voiceStylePaths[0]);
|
||
const firstStyle = await firstResponse.json();
|
||
|
||
const ttlDims = firstStyle.style_ttl.dims;
|
||
const dpDims = firstStyle.style_dp.dims;
|
||
|
||
const ttlDim1 = ttlDims[1];
|
||
const ttlDim2 = ttlDims[2];
|
||
const dpDim1 = dpDims[1];
|
||
const dpDim2 = dpDims[2];
|
||
|
||
// Pre-allocate arrays with full batch size
|
||
const ttlSize = bsz * ttlDim1 * ttlDim2;
|
||
const dpSize = bsz * dpDim1 * dpDim2;
|
||
const ttlFlat = new Float32Array(ttlSize);
|
||
const dpFlat = new Float32Array(dpSize);
|
||
|
||
// Fill in the data
|
||
for (let i = 0; i < bsz; i++) {
|
||
const response = await fetch(voiceStylePaths[i]);
|
||
const voiceStyle = await response.json();
|
||
|
||
// Flatten TTL data
|
||
const ttlData = voiceStyle.style_ttl.data.flat(Infinity);
|
||
const ttlOffset = i * ttlDim1 * ttlDim2;
|
||
ttlFlat.set(ttlData, ttlOffset);
|
||
|
||
// Flatten DP data
|
||
const dpData = voiceStyle.style_dp.data.flat(Infinity);
|
||
const dpOffset = i * dpDim1 * dpDim2;
|
||
dpFlat.set(dpData, dpOffset);
|
||
}
|
||
|
||
const ttlShape = [bsz, ttlDim1, ttlDim2];
|
||
const dpShape = [bsz, dpDim1, dpDim2];
|
||
|
||
const ttlTensor = new ort.Tensor('float32', ttlFlat, ttlShape);
|
||
const dpTensor = new ort.Tensor('float32', dpFlat, dpShape);
|
||
|
||
if (verbose) {
|
||
console.log(`Loaded ${bsz} voice styles`);
|
||
}
|
||
|
||
return new Style(ttlTensor, dpTensor);
|
||
}
|
||
|
||
/**
|
||
* Load configuration from JSON
|
||
*/
|
||
export async function loadCfgs(onnxDir) {
|
||
const response = await fetch(`${onnxDir}/tts.json`);
|
||
const cfgs = await response.json();
|
||
return cfgs;
|
||
}
|
||
|
||
/**
|
||
* Load text processor
|
||
*/
|
||
export async function loadTextProcessor(onnxDir) {
|
||
const response = await fetch(`${onnxDir}/unicode_indexer.json`);
|
||
const indexer = await response.json();
|
||
return new UnicodeProcessor(indexer);
|
||
}
|
||
|
||
/**
|
||
* Load ONNX model
|
||
*/
|
||
export async function loadOnnx(onnxPath, options) {
|
||
const session = await ort.InferenceSession.create(onnxPath, options);
|
||
return session;
|
||
}
|
||
|
||
/**
|
||
* Load all TTS components
|
||
*/
|
||
export async function loadTextToSpeech(onnxDir, sessionOptions = {}, progressCallback = null) {
|
||
console.log('Using WebAssembly/WebGPU for inference');
|
||
|
||
const cfgs = await loadCfgs(onnxDir);
|
||
|
||
const dpPath = `${onnxDir}/duration_predictor.onnx`;
|
||
const textEncPath = `${onnxDir}/text_encoder.onnx`;
|
||
const vectorEstPath = `${onnxDir}/vector_estimator.onnx`;
|
||
const vocoderPath = `${onnxDir}/vocoder.onnx`;
|
||
|
||
const modelPaths = [
|
||
{ name: 'Duration Predictor', path: dpPath },
|
||
{ name: 'Text Encoder', path: textEncPath },
|
||
{ name: 'Vector Estimator', path: vectorEstPath },
|
||
{ name: 'Vocoder', path: vocoderPath }
|
||
];
|
||
|
||
const sessions = [];
|
||
for (let i = 0; i < modelPaths.length; i++) {
|
||
if (progressCallback) {
|
||
progressCallback(modelPaths[i].name, i + 1, modelPaths.length);
|
||
}
|
||
const session = await loadOnnx(modelPaths[i].path, sessionOptions);
|
||
sessions.push(session);
|
||
}
|
||
|
||
const [dpOrt, textEncOrt, vectorEstOrt, vocoderOrt] = sessions;
|
||
|
||
const textProcessor = await loadTextProcessor(onnxDir);
|
||
const textToSpeech = new TextToSpeech(cfgs, textProcessor, dpOrt, textEncOrt, vectorEstOrt, vocoderOrt);
|
||
|
||
return { textToSpeech, cfgs };
|
||
}
|
||
|
||
/**
|
||
* Chunk text into manageable segments
|
||
*/
|
||
function chunkText(text, maxLen = 300) {
|
||
if (typeof text !== 'string') {
|
||
throw new Error(`chunkText expects a string, got ${typeof text}`);
|
||
}
|
||
|
||
// Split by paragraph (two or more newlines)
|
||
const paragraphs = text.trim().split(/\n\s*\n+/).filter(p => p.trim());
|
||
|
||
const chunks = [];
|
||
|
||
for (let paragraph of paragraphs) {
|
||
paragraph = paragraph.trim();
|
||
if (!paragraph) continue;
|
||
|
||
// Split by sentence boundaries (period, question mark, exclamation mark followed by space)
|
||
// But exclude common abbreviations like Mr., Mrs., Dr., etc. and single capital letters like F.
|
||
const sentences = paragraph.split(/(?<!Mr\.|Mrs\.|Ms\.|Dr\.|Prof\.|Sr\.|Jr\.|Ph\.D\.|etc\.|e\.g\.|i\.e\.|vs\.|Inc\.|Ltd\.|Co\.|Corp\.|St\.|Ave\.|Blvd\.)(?<!\b[A-Z]\.)(?<=[.!?])\s+/);
|
||
|
||
let currentChunk = "";
|
||
|
||
for (let sentence of sentences) {
|
||
if (currentChunk.length + sentence.length + 1 <= maxLen) {
|
||
currentChunk += (currentChunk ? " " : "") + sentence;
|
||
} else {
|
||
if (currentChunk) {
|
||
chunks.push(currentChunk.trim());
|
||
}
|
||
currentChunk = sentence;
|
||
}
|
||
}
|
||
|
||
if (currentChunk) {
|
||
chunks.push(currentChunk.trim());
|
||
}
|
||
}
|
||
|
||
return chunks;
|
||
}
|
||
|
||
/**
|
||
* Write WAV file to ArrayBuffer
|
||
*/
|
||
export function writeWavFile(audioData, sampleRate) {
|
||
const numChannels = 1;
|
||
const bitsPerSample = 16;
|
||
const byteRate = sampleRate * numChannels * bitsPerSample / 8;
|
||
const blockAlign = numChannels * bitsPerSample / 8;
|
||
const dataSize = audioData.length * 2;
|
||
|
||
// Create ArrayBuffer
|
||
const buffer = new ArrayBuffer(44 + dataSize);
|
||
const view = new DataView(buffer);
|
||
|
||
// Write WAV header
|
||
const writeString = (offset, string) => {
|
||
for (let i = 0; i < string.length; i++) {
|
||
view.setUint8(offset + i, string.charCodeAt(i));
|
||
}
|
||
};
|
||
|
||
writeString(0, 'RIFF');
|
||
view.setUint32(4, 36 + dataSize, true);
|
||
writeString(8, 'WAVE');
|
||
writeString(12, 'fmt ');
|
||
view.setUint32(16, 16, true);
|
||
view.setUint16(20, 1, true); // PCM
|
||
view.setUint16(22, numChannels, true);
|
||
view.setUint32(24, sampleRate, true);
|
||
view.setUint32(28, byteRate, true);
|
||
view.setUint16(32, blockAlign, true);
|
||
view.setUint16(34, bitsPerSample, true);
|
||
writeString(36, 'data');
|
||
view.setUint32(40, dataSize, true);
|
||
|
||
// Write audio data
|
||
const int16Data = new Int16Array(audioData.length);
|
||
for (let i = 0; i < audioData.length; i++) {
|
||
const clamped = Math.max(-1.0, Math.min(1.0, audioData[i]));
|
||
int16Data[i] = Math.floor(clamped * 32767);
|
||
}
|
||
|
||
const dataView = new Uint8Array(buffer, 44);
|
||
dataView.set(new Uint8Array(int16Data.buffer));
|
||
|
||
return buffer;
|
||
}
|