matcha-tts-pl / cleaner.js
mcPear's picture
Questions: '?' glued to the word, as in v2's training text (a space made wh-questions rise)
2125ef1 verified
Raw History Blame Contribute Delete
5.37 kB
// Browser port of matcha_patch/polish_cleaners.py + polish_mixed_cleaners: Polish text normalisation, espeak-ng (WASM) IPA,
// English words in {braces} phonemized with the en-us voice, symbol ids. Must match the Python cleaners character for character.
import { initialize, setVoice, getPhonemes } from "https://cdn.jsdelivr.net/npm/espeak-phonemizer@0.1.2/src/espeak.mjs";
import { SYMBOL_TO_ID } from "./symbols.js?v=20261005c";
const PL_WORD = "0-9a-ząćęłńóśźż";
const esc = (s) => s.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
function abbr(k, v) {
if (k === "r." || k === "w.") return [new RegExp(`(?<=[0-9ivx])\\s?${esc(k)}`, "g"), ` ${v}`];
const tail = k.endsWith(".") ? "" : `(?![${PL_WORD}])`;
return [new RegExp(`(?<![${PL_WORD}])${esc(k)}${tail}`, "g"), v];
}
const ABBREV = [["np.", "na przykład"], ["tzn.", "to znaczy"], ["tj.", "to jest"], ["itd.", "i tak dalej"], ["itp.", "i tym podobne"],
["m.in.", "między innymi"], ["dr", "doktor"], ["prof.", "profesor"], ["inż.", "inżynier"], ["mgr", "magister"], ["ul.", "ulica"],
["al.", "aleja"], ["r.", "roku"], ["w.", "wieku"], ["godz.", "godzina"], ["min.", "minut"], ["św.", "święty"], ["nr", "numer"],
["tel.", "telefon"], ["zł", "złotych"]].map(([k, v]) => abbr(k, v));
export function normalizeText(text) {
let t = text.toLowerCase();
t = t.replace(/\s*(?:—|–| - )\s*/g, ", ").replace(/…/g, "...");
t = t.replace(/[„”“"«»]/g, "");
for (const [rx, rep] of ABBREV) t = t.replace(rx, rep);
return t;
}
let ready = null, voice = "";
export function initPhonemizer() {
if (!ready) ready = (async () => { await initialize("https://cdn.jsdelivr.net/npm/espeak-phonemizer@0.1.2/dist"); await setVoice("pl"); voice = "pl"; })();
return ready;
}
async function useVoice(v) { if (voice !== v) { await setVoice(v); voice = v; } }
// phonemizer (python) splits on punctuation, phonemizes each chunk and glues the punctuation back
const PUNCT = /([;:,.!?]+)/;
function espeakChunks(t) {
let out = "", prev = "";
for (const part of t.split(PUNCT)) {
if (!part) continue;
if (/^[;:,.!?]+$/.test(part)) { out += (/\s$/.test(prev) && out ? " " : "") + part; prev = part; continue; } // keep "word ?" like phonemizer
const afterPunct = /^[;:,.!?]+$/.test(prev), glued = afterPunct && !/^\s/.test(part); // "Co?Jak": no space, like phonemizer
prev = part;
const chunk = part.trim();
if (!chunk) continue;
const ph = getPhonemes(chunk).map(c => (c.phonemes || "").trim()).filter(Boolean).join(" ");
if (!ph) continue;
out += (out && !out.endsWith(" ") && !glued ? " " : "") + ph;
}
return out.replace(/\s+/g, " ").trim();
}
// this espeak-ng WASM build writes "iʲ" before a vowel ("znajomymiʲ albo"); the training phonemizer (espeak-ng 1.50, Linux)
// never does ("iʲ" does not occur in the training data at all) -> normalise, so the model gets what it was trained on
const fixPl = (ph) => ph.replace(/iʲ/g, "i");
async function plPhon(text) { await useVoice("pl"); const t = normalizeText(text); return t.trim() ? fixPl(espeakChunks(t)) : ""; }
async function enPhon(text) { await useVoice("en-us"); return espeakChunks(text).replace(/ʲ/g, ""); } // same WASM quirk; en-us never has ʲ
const VOICELESS = ["p", "t", "k", "f", "s", "ʃ", "θ", "h", "tʃ", "ˈp", "ˈt", "ˈk", "ˈf", "ˈs", "ˈʃ", "ˈθ", "ˈh", "ˌp", "ˌt", "ˌk", "ˌf", "ˌs", "ˌh"];
function oneLetter(w, nextPh) {
const vl = VOICELESS.some(p => nextPh.startsWith(p));
return { w: vl ? "f" : "v", z: vl ? "s" : "z", i: "i", a: "a", o: "ɔ", u: "u" }[w];
}
const unstress = (ph) => { const i = ph.indexOf(" "); const head = i < 0 ? ph : ph.slice(0, i); return head.replace(/[ˈˌ]/g, "") + (i < 0 ? "" : ph.slice(i)); };
export async function textToIpa(text) {
await initPhonemizer();
text = text.replace(/\s+\?/g, "?"); // v2 was trained on VoxCPM2 texts with '?' glued to the word: "kuchni ?" -> "kuchni?" (a space makes wh-questions rise)
if (!text.includes("{")) return plPhon(text);
let out = "", pos = 0;
for (const m of text.matchAll(/\{([^{}]+)\}/g)) {
let pl = text.slice(pos, m.index), prep = "";
const mm = pl.match(/(?:^|\s)([wzaiouWZAIOU])\s*$/);
if (mm) { prep = mm[1].toLowerCase(); pl = pl.slice(0, mm.index + mm[0].indexOf(mm[1])); } // = Python pl[:mm.start(1)]
if (pl.trim()) {
const glue = !/^\s/.test(pl) && out !== "";
let ph = await plPhon(pl);
if (glue) ph = unstress(ph);
out += (glue ? "" : " ") + ph;
}
const en = await enPhon(m[1]);
if (prep) out += " " + oneLetter(prep, en);
const glue = m.index > 0 && !/\s/.test(text[m.index - 1]) && out !== "";
out += (glue ? "" : " ") + en;
pos = m.index + m[0].length;
}
const tail = text.slice(pos);
if (tail.trim()) {
const glue = !/^\s/.test(tail);
let ph = await plPhon(tail);
if (glue) ph = unstress(ph);
out += (glue ? "" : " ") + ph;
}
await useVoice("pl");
return out.replace(/\s+/g, " ").trim();
}
export function ipaToIds(ipa) {
const ids = [];
const unknown = new Set();
for (const ch of ipa) {
const id = SYMBOL_TO_ID[ch];
if (id === undefined) unknown.add(ch); else ids.push(id);
}
const seq = new Array(ids.length * 2 + 1).fill(0);
ids.forEach((v, i) => { seq[2 * i + 1] = v; });
return { seq, unknown: [...unknown] };
}