Pocket TTS

Local text to speech for @micdrop/server, running Pocket TTS in your Node process through the sherpa-onnx addon and ONNX Runtime.

Pocket TTS is a 100 million parameter model from Kyutai, designed for the CPU, that takes its voice from a few seconds of reference audio instead of from a catalog. It generates about five times faster than real time and emits its audio while it is still writing the rest, which is what makes it the local voice with the shortest delay before the first word.

Installation

Install the package:

npm install @micdrop/pocket-tts

Download the weights, converted to ONNX by the sherpa-onnx project:

curl -LO https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/sherpa-onnx-pocket-tts-int8-2026-01-26.tar.bz2
tar xf sherpa-onnx-pocket-tts-int8-2026-01-26.tar.bz2
rm sherpa-onnx-pocket-tts-int8-2026-01-26.tar.bz2

The archive weighs 98 MB and takes 190 MB on disk. A full precision archive, sherpa-onnx-pocket-tts-2026-01-26.tar.bz2, sits next to it and is read the same way: the weights are looked up by name in the folder, quantized first when both are there.

Usage

import { PocketTTS } from '@micdrop/pocket-tts'
import { MicdropServer } from '@micdrop/server'

const tts = new PocketTTS({
  modelDir: './sherpa-onnx-pocket-tts-int8-2026-01-26',
  voice: 'bria',
})

// Use with MicdropServer
new MicdropServer(socket, {
  tts,
  // ... other options
})

Options

Voices

The voice comes from a recording rather than from a catalog: a few seconds of someone speaking is enough for the model to answer in that voice, with no transcript of the sample and no training step. The archive ships three of them in its test_wavs folder, reachable by name through BUNDLED_VOICES, and any wav file works in their place.

import { BUNDLED_VOICES, PocketTTS } from '@micdrop/pocket-tts'

const tts = new PocketTTS({
  modelDir: './sherpa-onnx-pocket-tts-int8-2026-01-26',
  voice: './voices/my-voice.wav',
})

bria is a woman reading calmly for forty seconds, loona is a one second clip, which is about the shortest a reference can be, and frenchAccent is a French speaker: the words come from the text and stay English, the timbre and the accent come from the recording.

Only the first referenceLength seconds are read, so a long recording costs nothing extra. The quality of the sample is reproduced along with the voice, which makes a clean recording worth more than a long one.

Languages

Pocket TTS speaks English here. Kyutai also publishes French, German, Spanish, Italian and Portuguese checkpoints, and none of them is converted to ONNX yet, so the English model is the one the addon loads. Handing it French produces fluent nonsense rather than an accent, since it reads every word with English phonemes. Use Piper for the other languages.

Latency

Measured on an Apple M2, on CPU, with the model already loaded:

The answer is cut into sentences, and each sentence is handed over in chunks as the model writes it, so the user hears its opening words while the model is still generating its end. An interruption stops the generation between two chunks rather than at the end of the sentence, which is what keeps a barge-in immediate.

Memory

The model takes about 350 MB once loaded and around 600 MB while generating, which is more than Kokoro at 80 MB and small next to the language model it runs beside. One instance is loaded per configuration and shared by every call, so a second conversation costs nothing more.

Documentation

Read the guide on running Micdrop with local models to see how this fits with a local LLM and a local transcription.

License

Pocket TTS is released under the MIT license, and the sherpa-onnx addon under Apache 2.0. The reference recordings shipped with the archive carry their own licenses, listed on the Kyutai voice repository.