kitten-tts-js

kitten-tts-js

JavaScript/TypeScript port of KittenTTS — ultra-lightweight neural TTS via ONNX. Works in Node.js, browser (WebAssembly), and any JS environment. Zero Python dependency.

Live Demo → · npm → · GitHub →

Based on KittenTTS by KittenML / Stellon Labs — original Python library: github.com/KittenML/KittenTTS — original models & voices: huggingface.co/KittenML

All credit for the models, architecture, and voice embeddings goes to them. Licensed under Apache 2.0. See NOTICE for full attribution.

Disclaimer: This is an unofficial community port made by a hobbyist who needed KittenTTS in JavaScript. It is not affiliated with, endorsed by, or supported by KittenML or Stellon Labs.

Features

• Ultra-lightweight — nano model is ~25 MB
• Runs anywhere — Node.js (CPU), browser (WASM), Cloudflare Workers
• 8 voices — Bella, Luna, Rosie, Kiki, Leo, Jasper, Bruno, Hugo
• StyleTTS2-based ONNX models from HuggingFace
• Streaming support — sentence-by-sentence async generator
• TypeScript declarations included
• Automatic caching — ~/.cache/kitten-tts/ in Node, Cache API in browser

Install

npm install kitten-tts-js

Quick Start

Node.js

import { KittenTTS } from 'kitten-tts-js';

const tts = await KittenTTS.from_pretrained('KittenML/kitten-tts-nano-0.8');

console.log(tts.list_voices());
// → ['Bella', 'Jasper', 'Luna', 'Bruno', 'Rosie', 'Hugo', 'Kiki', 'Leo']

const audio = await tts.generate('Hello from KittenTTS!', { voice: 'Bella' });
await audio.save('output.wav');

Browser (inline)

<script type="module">
  import { KittenTTS } from 'https://esm.sh/kitten-tts-js';

  const tts = await KittenTTS.from_pretrained('KittenML/kitten-tts-nano-0.8');
  const audio = await tts.generate('Hello!', { voice: 'Luna' });

  const audioCtx = new AudioContext();
  const source = audioCtx.createBufferSource();
  source.buffer = audio.toAudioBuffer(audioCtx);
  source.connect(audioCtx.destination);
  source.start();
</script>

Browser (Web Worker — recommended for production)

Running inference in a Worker keeps the UI thread responsive during the ~5–10 s model load and synthesis.

worker.js

import { KittenTTS } from 'https://esm.sh/kitten-tts-js';
let tts;

self.onmessage = async ({ data }) => {
  if (data.type === 'load') {
    tts = await KittenTTS.from_pretrained(data.modelId);
    self.postMessage({ type: 'ready' });
  }
  if (data.type === 'generate') {
    const audio = await tts.generate(data.text, data.opts);
    const buf = new Float32Array(audio.data);
    self.postMessage({ type: 'audio', buf, sampleRate: audio.sampling_rate }, [buf.buffer]);
  }
};

main.js

const worker = new Worker('./worker.js', { type: 'module' });
worker.postMessage({ type: 'load', modelId: 'KittenML/kitten-tts-nano-0.8' });

worker.onmessage = ({ data }) => {
  if (data.type === 'ready') console.log('Model loaded!');
  if (data.type === 'audio') playFloat32(data.buf, data.sampleRate);
};

worker.postMessage({ type: 'generate', text: 'Hello world!', opts: { voice: 'Bella' } });

function playFloat32(buf, sampleRate) {
  const audioCtx = new AudioContext({ sampleRate });
  const ab = audioCtx.createBuffer(1, buf.length, sampleRate);
  ab.copyToChannel(buf, 0);
  const src = audioCtx.createBufferSource();
  src.buffer = ab;
  src.connect(audioCtx.destination);
  src.start();
}

Streaming (sentence-by-sentence)

let i = 0;
for await (const { text, audio } of tts.stream(longText, { voice: 'Leo' })) {
  console.log(`Chunk: "${text}" → ${audio.duration.toFixed(1)}s`);
  await audio.save(`chunk-${i++}.wav`);
}

API

KittenTTS.from_pretrained(modelId?, opts?)

Param	Type	Default	Description
modelId	string	'KittenML/kitten-tts-nano-0.8'	HuggingFace repo ID
opts.cacheDir	string	~/.cache/kitten-tts	Override cache dir (Node)

tts.generate(text, opts?)

Returns Promise<RawAudio>.

Opt	Default	Description
voice	'Leo'	Voice name (see table below)
speed	1.0	Speed multiplier (0.5–2.0)
clean	true	Run text preprocessor (numbers, currency, etc.)

tts.stream(text, opts?)

Returns AsyncGenerator<{ text: string, audio: RawAudio }> — one chunk per sentence.

tts.list_voices()

Returns string[] of available friendly voice names.

tts.release()

Releases the underlying ONNX session to free WebAssembly memory. Useful when switching models in the browser.

RawAudio

Member	Description
.data	Float32Array — raw PCM mono
.sampling_rate	24000
.duration	Duration in seconds
.toWav()	ArrayBuffer — 16-bit PCM WAV
.save(path)	Write WAV file (Node.js)
.toBlob()	Blob for browser download/playback
.toAudioBuffer(ctx)	Web Audio AudioBuffer

Available Models

Model ID	Size	Speed	Quality
KittenML/kitten-tts-nano-0.8	~25 MB	★★★	★★☆
KittenML/kitten-tts-micro-0.8	~40 MB	★★☆	★★★
KittenML/kitten-tts-mini-0.8	~80 MB	★☆☆	★★★

Available Voices

Friendly Name	Gender
Bella	Female
Jasper	Male
Luna	Female
Bruno	Male
Rosie	Female
Hugo	Male
Kiki	Female
Leo	Male

Development

git clone https://github.com/Algiras/kitten-tts-js.git
cd kitten-tts-js
npm install
npm test              # run unit tests
npm run build:pages   # build browser bundle → docs/

Architecture

src/
├── kitten-tts.js    Main class: from_pretrained, generate, stream
├── preprocess.js    Number/currency/time text normalization
├── text-cleaner.js  Phoneme → token IDs (IPA symbol table)
├── phonemizer.js    eSpeak-NG WASM phonemization
├── npz-loader.js    NumPy .npz binary parser
├── model-loader.js  HuggingFace Hub download + caching
├── audio.js         RawAudio class + WAV encoder
└── index.js         Public API re-exports

License

Apache 2.0 — see NOTICE for attribution to the original KittenTTS by KittenML / Stellon Labs.