Skip to content

Browser TTS workspace

Pocket TTS

Run Pocket TTS locally with 8 built-in voices and zero-shot voice cloning across 5 languages. Free browser synthesis, no signup, WAV & MP3 export.

  • Private generation
  • WAV + MP3 export
  • 8 voices · cloning
MB model
146
int8 · WASM
MB model
voices
8
built-in + cloning
voices
languages
5
en de it pt es
languages
runtime
WASM
+WebGPU option
runtime

Audio generation uses WebGPU/WASM. Desktop Chrome or Edge gives the most reliable result.

About Pocket TTS

Pocket TTS is Kyutai's 100M-parameter text-to-speech model, designed to run on modest hardware. By default this integration executes the full model in the browser through ONNX Runtime Web on CPU WebAssembly — with an optional WebGPU backend toggle, no API key, and no cloud synthesis.

The English bundle ships with 8 built-in voices. You can also clone your own voice two ways in WASM mode: record it directly with your microphone, or upload a short, clean audio clip. In both cases the mimi voice encoder runs locally to build a custom voice, so your audio never leaves the device. Five language bundles (English, German, Italian, Portuguese, Spanish) are fetched on demand and cached after the first download; each non-English bundle is roughly 146MB, while the English bundle is about 198MB including its built-in voices file.

Generation streams in chunks, is buffered, and then plays through the shared result panel with WAV or MP3 export. Because default inference runs on CPU WebAssembly, actual speed depends on the device, browser, and text length — measure on the intended hardware before planning bulk work.

Compare engines: Kokoro TTS (54 voices · q4 or fp32) · Kitten TTS (8 voices · Lightest) · Piper TTS (25 voices · Fastest CPU) · Supertonic TTS (31 languages · Local)

Advertisement

Adsterra privacy

Pocket Network Boundary and Voice Cloning Review

Pocket TTS runs inference fully in the browser via ONNX Runtime Web on WebAssembly. The language bundles and the per-language voices file are downloaded from Hugging Face at runtime and cached in the browser; after those assets are available, synthesis and voice cloning run on the device. Model delivery and ordinary site analytics are separate network activity, and cloned reference audio is processed locally by the mimi encoder.

Cloned output quality depends heavily on the reference clip: keep it short, clean, single-speaker, and free of background noise, and compare the clone on representative text before a long export. The eight built-in voice names are catalog labels from the model's conditioning states rather than verified identities or a quality ranking. WAV preserves the browser-generated PCM for editing; MP3 is smaller but lossy.

When Pocket TTS Is the Appropriate Engine

Choose Pocket when a CPU-friendly WebAssembly path and on-device zero-shot cloning matter more than a large voice catalog. It is a compact 100M-parameter model with 8 built-in voices, local cloning from a short sample, and five language bundles that load on demand. Kitten is lighter, Kokoro has far more voices and languages, Piper is fast on many CPUs, and Supertonic covers far more languages — Pocket's edge is compact local synthesis with built-in voice cloning.

Pocket is not a hosted API, a guarantee of clone fidelity, or a promise of equal speed across devices: default WASM inference is CPU-bound (with an optional WebGPU path for built-in voices), so generation time varies with the processor, browser, memory, and text length. Review the final audio on the intended device, and retain the reference clip, voice, language, and settings with an important production file.

Getting Started with Pocket TTS

  1. 1

    Load the Bundle

    The first load downloads and caches the selected language bundle from Hugging Face. Non-English bundles are roughly 146MB; the English bundle is about 198MB including its voices file. Later visits reuse the browser cache.

  2. 2

    Pick a Voice, Record, or Upload

    Choose one of the 8 built-in voices, record your own voice with the microphone, or upload a short clean audio clip (up to 10 seconds) to create a custom cloned voice locally. The clone is ready for the next generation.

  3. 3

    Enter Your Text

    Type or paste up to 50,000 characters in the selected language. Natural punctuation keeps chunking and pacing clean across the streaming generation.

  4. 4

    Generate, Clone, Export

    Generate speech locally, preview it in the shared result panel, then download WAV for editing or MP3 for a smaller review file. A cloned voice stays active until you switch voices.

Tips for Pocket TTS

  • 1

    Record or upload a clean, short reference. Cloning quality depends on the reference clip: a few seconds of clean, single-speaker audio with minimal background noise gives the best results. The microphone records directly; for uploaded files, longer or noisy samples can reduce intelligibility.

  • 2

    Treat voice labels as catalog labels. The 8 built-in voice names come from the model's conditioning states. They are catalog labels, not verified identities of named people or a quality ranking.

  • 3

    Use WASM for cloning and expect CPU-bound speed. Pocket defaults to WebAssembly, which is required for custom voice cloning and Voice Library references. Generation speed varies with device, browser, memory, and text length, so test on target hardware before batch work.

  • 4

    Know the network boundary. Model bundles and the voices file are downloaded from Hugging Face. After the required assets are cached, inference and cloning run on the device; site analytics remain separate network activity.

Frequently Asked Questions about Pocket TTS

How does zero-shot voice cloning work in Pocket TTS?

Upload a clean 3–10 second audio clip or record directly from your microphone in WASM mode. The local mimi voice encoder extracts speaker characteristics inside your browser without uploading your reference recording.

Should I choose WASM or WebGPU in Pocket TTS?

Keep WASM selected when using microphone recording, uploaded audio cloning, or Voice Library references. WebGPU is available as an experimental backend option for built-in presets on supported browsers.

Which languages does Pocket TTS support?

Pocket TTS provides on-demand language bundles for English (~198MB including built-in voices), German, Italian, Portuguese, and Spanish (~146MB each), cached in browser storage after the first load.

Are cloned custom voices uploaded or shared?

No. Encoded custom voice prompts and generated audio stay in your browser’s local storage (IndexedDB) until you delete them, and exported speech downloads directly as WAV or MP3.

Upstream Model & Runtime Sources