verb · → audio

speak

Turn text, or a live stream of text, into speech.

Speak and play#

import { speak } from 'edgewise';

// browser: start playing with the first sentence
await speak({ model: 'voice:default', voice: 'af_heart', input: 'Hello from your device.' }).play();

// anywhere: get the audio
const audio = await speak({ model: 'voice:default', input: 'Hello.' });
audio.samples;    // Float32Array, mono
audio.sampleRate; // 24000
audio.duration;   // seconds
audio.toWav();    // Uint8Array, 16-bit PCM WAV

speak needs the optional phonemizer package for Kokoro. It bundles espeak-ng, which is GPL-3.0; check that it suits your app.

Stream text in#

input can be any async iterable of strings, such as a running generate(). Speech is made one sentence at a time, as soon as each sentence is complete:

const reply = generate({ model: 'text:default', input: question });
const run = speak({ model: 'voice:default', input: reply });

for await (const chunk of run) console.log(chunk.text, chunk.samples.length);

.play() schedules those chunks back to back without gaps.

Voices#

import { listVoices } from 'edgewise';

const voices = await listVoices('voice:default'); // [{ id: 'af_heart', name: 'Heart', lang: 'en-US', gender: 'female', grade: 'A' }, …]

await speak({ model: 'voice:default', voice: 'bm_george', input: 'Good evening.' });
await speak({ model: 'voice:default', voice: { af_heart: 0.7, af_bella: 0.3 }, input: 'A blended voice.' }); // weighted blend
await speak({ model: 'voice:default', speed: 1.2, input: 'A little faster.' }); // 0.5 to 2, Kokoro only

Kokoro has 28 English voices: American (a…) and British (b…), female (?f_) and male (?m_).

Voice cloning#

chatterbox-turbo (alias voice:clone) speaks in any voice from 3 to 10 seconds of reference audio. Cloning requires you to confirm the speaker agreed:

const reference = await fetch('/ana-sample.wav').then((r) => r.blob()); // one speaker, clear speech

await speak({
  model: 'voice:clone',
  allowPreview: true,
  voice: { reference, consent: { attested: true, by: 'Ana, signed release 2026-09-01' } },
  input: 'Hi, this is Ana. Well, a copy of her voice.',
}).play();

Encoding the reference takes a moment, so reuse it:

import { cloneVoice } from 'edgewise';

const ana = await cloneVoice({ reference, consent: { attested: true }, saveAs: 'ana', allowPreview: true });
await speak({ model: 'voice:clone', voice: ana, input: 'First line.', allowPreview: true });
await speak({ model: 'voice:clone', voice: 'saved:ana', input: 'Tomorrow, after a reload.', allowPreview: true });

saveAs keeps the voice (the speaker conditioning, not the recording) in the Origin Private File System in browsers and in the cache folder on servers. exaggeration (0 to 2, default 0.5) makes the delivery more or less expressive; speed is not supported.

Chatterbox Turbo is English only and downloads about 720 MB. Its 4-bit graphs need WebGPU in browsers; on servers it runs on the CPU (on two CPU cores, about four to five seconds of work per second of speech).

Careful

A cloned voice can be used to impersonate someone. Clone only voices you have permission to use, tell listeners when speech is synthetic, and use configure({ speak: { onSynthesize } }) to log what your app generates.

Models#

Model	Accepts	Download	Status	Notes
`kokoro-82m`	text	88 MB	stable	`voice:default` · High-quality English speech with 28 voices.
`chatterbox-turbo`	text, audio	684 MB	preview	`voice:clone` · Zero-shot voice cloning from 3 to 10 seconds of reference audio. English. Needs WebGPU in browsers.