@revolab/revolab-edge

Browser-native neural Malay text-to-speech. Malay text in, waveform out, entirely on-device — after the assets load you can go offline.

Install

npm install @revolab/revolab-edge

Quick start

import { RevolabEdge, playAudio } from '@revolab/revolab-edge';

const tts = await RevolabEdge.load({ assetBase: '/assets/revolab-edge/' });
const { samples, sampleRate } = await tts.synthesize(
  'Selamat pagi! Saya suara Revolab dalam pelayar anda.',
  { voice: 'sarah' },
);
playAudio({ samples, sampleRate });

assetBase is where you host this package's files (wasm, g2p.json, voices/). Point it at a copy on your own server, or use a CDN: https://unpkg.com/@revolab/revolab-edge/.

Voices

Three Malay speakers:

| Voice | Key | Language | Parameters | Weights | |---|---|---|---|---| | angellyn | angellyn | Malay (ms) | 1,394,445 | ~1.4 MB | | sarah | sarah | Malay (ms) | 1,394,445 | ~1.4 MB | | paan | paan | Malay (ms) | 1,394,445 | ~1.4 MB |

All voices output 22.05 kHz mono PCM. Weights are int8-quantized (~1.4 MB per voice), fetched lazily on first use and cached for the page's life — switching voices after the first load is instant. Pass { quantized: false } to synthesize for the fp32 weights (~5.6 MB) instead.

await tts.loadVoice('paan');                    // prefetch (e.g. on hover)
const r = await tts.synthesize(text, { voice: 'paan' });

What you can do with it

Voice picker UI — KNOWN_VOICES has key/label/language/flag for building selectors:

import { KNOWN_VOICES } from '@revolab/revolab-edge';
KNOWN_VOICES.forEach(v => picker.add({ value: v.key, label: `${v.flag} ${v.label}` }));

Speaking-rate control — lengthScale slows (larger) or speeds up (smaller) speech:

await tts.synthesize(text, { voice: 'sarah', lengthScale: 1.3 });  // slower
await tts.synthesize(text, { voice: 'sarah', lengthScale: 0.85 }); // faster

Live latency meter — elapsedMs is the honest on-device wall time for the synth call; show real-time factor to your users:

const { samples, elapsedMs } = await tts.synthesize(text, { voice: 'paan' });
const rtf = (elapsedMs / 1000) / (samples.length / 22050);
console.log(`RTF ${rtf.toFixed(2)} on this machine`);

Waveform visualization / custom playback — samples is a plain Float32Array; you don't have to use playAudio:

const r = await tts.synthesize(text, { voice: 'angellyn' });
drawWaveform(r.samples);                        // your canvas code
const buf = wavEncode(r.samples, r.sampleRate); // your encoder
downloadBlob(buf, 'speech.wav');

Reuse one AudioContext — pass yours to avoid leaking a new one per call:

const ctx = new AudioContext();
playAudio(r, { audioContext: ctx });

Long text — pass a bigger maxSeconds (output buffer cap, default 30):

await tts.synthesize(paragraph, { voice: 'sarah', maxSeconds: 120 });

Text normalization — numbers, currency, dates, and abbreviations should be normalized to words before synthesis (the voices were trained on normalized text). Use @revolab/revonorm (Rust/Wasm, ms/id/en/zh support):

import { normalize_malay } from '@revolab/revonorm/revonorm_core.js';

const text = normalize_malay('Harga RM10.50 sahaja');
// 'Harga sepuluh ringgit lima puluh sen sahaja'
await tts.synthesize(text, { voice: 'sarah' });

revonorm is optional and loaded independently — revolab-edge accepts any pre-normalized string.

API

• RevolabEdge.load({ assetBase }) → Promise<RevolabEdge> — instantiate the wasm runtime, initialize the G2P. Call once per page.
• tts.synthesize(text, opts) → Promise<{samples, sampleRate, elapsedMs}> — Malay text to PCM. Opts: voice, voiceBase, lengthScale, maxSeconds. Voices lazy-load on first use.
• tts.loadVoice(key, opts) — prefetch + cache a voice's weights.
• playAudio(result, { audioContext }) — WebAudio playback; returns the started source node.
• KNOWN_VOICES — [{key, label, language, flag}] registry.

License

Proprietary — © Revolab.