d9e0f89b89
Some checks failed
standards workspaces build + test / jouleclaw-rs (push) Successful in 3m1s
standards workspaces build + test / aac-rs (push) Successful in 1m24s
standards workspaces build + test / joule-code-rs (push) Successful in 54s
standards workspaces build + test / joulecontract-rs (push) Successful in 59s
no-vanity-metrics / Scan committed Markdown for vanity metrics (push) Successful in 4s
standards workspaces build + test / map-rs (push) Successful in 1m8s
standards workspaces build + test / mesh-rs (push) Successful in 40s
standards workspaces build + test / eoc-rs (push) Successful in 1m49s
standards workspaces build + test / jcp-rs (push) Successful in 1m25s
standards workspaces build + test / openpay (push) Successful in 3m24s
standards workspaces build + test / proof-rs (push) Successful in 1m16s
standards workspaces build + test / sandbox-rs (push) Successful in 2m9s
standards workspaces build + test / smart-byte-rs (push) Successful in 2m36s
standards workspaces build + test / arl-rs (push) Has been cancelled
standards workspaces build + test / joule-ui-rs (push) Has been cancelled
standards workspaces build + test / jouledesk-rs (push) Has been cancelled
standards workspaces build + test / joulehook-rs (push) Has been cancelled
standards workspaces build + test / sense-rs (push) Has been cancelled
standards workspaces build + test / macOS energy backend (jcp-energy) (push) Has been cancelled
A better voice is a 100 MB download, and the person could not hear one before deciding. Making the samples showed something worse: some voices did not say the words at all. - Samples: each voice says a greeting and a request in its language (two or three seconds, MP3, about 10 KB). Settings > Better voices plays it with the sentence written beside it, before any download; checked by size and SHA-256 and played through the voices' own output. render_samples.mjs renders them with LAME; they are committed (Piper's synthesis is not repeatable), and build_voices.py publishes each beside the voices while it says the current text. - Intelligibility: intelligibility.py has every voice say seven everyday phrases, a speech recognizer (Whisper large-v3-turbo on MLX) writes down what it hears, and the scores are committed. Ratings now come from that evidence (RATING_RULE): under 0.80 the voice is withheld; under 0.90, or any phrase under 0.60, experimental; 0.90 and over clear; 0.95 and over with an upstream grade of A or B, natural. - Withheld, because they say other words: German (the only German voice, 0.14), Dutch MLS (0.27), Ukrainian (0.18) and both Japanese voices (0.63-0.69: they misread 外, 好き and 疲れた). Their languages say so plainly under "Why". Portuguese and Hindi are experimental; Spanish, Chinese, Italian, Flemish and Cori are clear; Heart, Emma and Siwis natural. 16 voices in 8 languages, 452 MB of packs. Co-Authored-By: Claude Opus 5.5 <[email protected]>
57 lines
2.9 KiB
JavaScript
57 lines
2.9 KiB
JavaScript
#!/usr/bin/env node
|
|
// Say each language's everyday phrases (phrases.json) with every voice in the
|
|
// built packs, into WAV files and a manifest, for intelligibility.py to
|
|
// transcribe. Nothing is sent anywhere.
|
|
//
|
|
// node tools/voices/intelligibility.mjs OUTDIR [--out PACKS] [voice-or-language ...]
|
|
import fs from 'node:fs';
|
|
import path from 'node:path';
|
|
import { fileURLToPath, pathToFileURL } from 'node:url';
|
|
|
|
const HERE = path.dirname(fileURLToPath(import.meta.url));
|
|
const argv = process.argv.slice(2);
|
|
const dir = path.resolve(argv.shift() || '');
|
|
let packs = process.env.AAC_VOICES_DIR || path.resolve(HERE, '../../../../../aac-voice-packs');
|
|
const only = [];
|
|
for (let i = 0; i < argv.length; i++) {
|
|
if (argv[i] === '--out') packs = path.resolve(argv[++i]);
|
|
else only.push(argv[i]);
|
|
}
|
|
const phrases = JSON.parse(fs.readFileSync(path.join(HERE, 'phrases.json'), 'utf8')).phrases;
|
|
const pointer = JSON.parse(fs.readFileSync(path.join(packs, 'index.json'), 'utf8'));
|
|
const cat = JSON.parse(fs.readFileSync(path.join(packs, pointer.current), 'utf8'));
|
|
const rt = Object.values(cat.runtimes)[0];
|
|
const { default: createModule } = await import(pathToFileURL(path.join(packs, rt.module)).href);
|
|
const m = await createModule({ wasmBinary: fs.readFileSync(path.join(packs, rt.wasm)), print: () => {}, printErr: () => {} });
|
|
const bytes = (f) => new Uint8Array(fs.readFileSync(path.join(packs, f.url)));
|
|
|
|
function wav(audio, sr) {
|
|
const b = Buffer.alloc(44 + audio.length * 2);
|
|
b.write('RIFF', 0); b.writeUInt32LE(36 + audio.length * 2, 4); b.write('WAVEfmt ', 8);
|
|
b.writeUInt32LE(16, 16); b.writeUInt16LE(1, 20); b.writeUInt16LE(1, 22); b.writeUInt32LE(sr, 24);
|
|
b.writeUInt32LE(sr * 2, 28); b.writeUInt16LE(2, 32); b.writeUInt16LE(16, 34); b.write('data', 36);
|
|
b.writeUInt32LE(audio.length * 2, 40);
|
|
for (let i = 0; i < audio.length; i++) b.writeInt16LE(Math.max(-32768, Math.min(32767, Math.round(audio[i] * 32767))), 44 + i * 2);
|
|
return b;
|
|
}
|
|
|
|
fs.mkdirSync(dir, { recursive: true });
|
|
const manifest = [];
|
|
for (const [code, lang] of Object.entries(cat.languages)) {
|
|
if (!phrases[code]) continue;
|
|
for (const pack of lang.packs) {
|
|
if (only.length && !only.includes(pack.id) && !only.includes(code)) continue;
|
|
const files = [...pack.requires.flatMap((c) => cat.components[c].files), ...pack.files];
|
|
const tts = new m.TextToSpeech(pack.lang_key, files.map((f) => f.name), files.map(bytes), Object.keys(pack.options), Object.values(pack.options));
|
|
phrases[code].forEach((text, i) => {
|
|
const r = tts.say(text);
|
|
const file = path.join(dir, `${pack.id}.${i}.wav`);
|
|
fs.writeFileSync(file, wav(r.audio, r.sampleRate));
|
|
manifest.push({ pack: pack.id, lang: code, text, file, seconds: Number((r.audio.length / r.sampleRate).toFixed(2)) });
|
|
});
|
|
tts.delete?.();
|
|
console.log(`said ${phrases[code].length} phrases with ${pack.id}`);
|
|
}
|
|
}
|
|
fs.writeFileSync(path.join(dir, 'manifest.json'), JSON.stringify(manifest, null, 1));
|
|
process.exit(0);
|