README.md

metadata

title: Pocket TTS ONNX Web Demo
emoji: 🌖
colorFrom: yellow
colorTo: pink
sdk: static
app_file: index.html
pinned: false
license: cc-by-4.0
short_description: Multilingual Pocket TTS voice cloning in the browser (CPU)
models:
  - KevinAHM/pocket-tts-onnx
custom_headers:
  cross-origin-embedder-policy: require-corp
  cross-origin-opener-policy: same-origin
  cross-origin-resource-policy: cross-origin

Pocket TTS Web Demo

Browser-only Pocket TTS inference with multilingual INT8 ONNX bundles and voice cloning.

Supported Bundles

• english_2026-04
• german
• italian
• portuguese
• spanish

The web demo intentionally skips the 24l variants and only ships the current April 2026 English checkpoint.

Features

• Multilingual bundle selector in the UI
• Built-in voices for every shipped language bundle
• Custom voice cloning from uploaded audio
• INT8 ONNX inference in the browser
• Streaming playback with low latency

Bundle Layout

Each language lives under onnx/<language>/ and includes:

• bundle.json
• tokenizer.model
• bos_before_voice.npy
• voices.bin
• mimi_encoder_int8.onnx
• text_conditioner_int8.onnx
• flow_lm_main_int8.onnx
• flow_lm_flow_int8.onnx
• mimi_decoder_int8.onnx

voices.bin is a local browser asset containing the compact built-in voice states for that language bundle.

Browser Requirements

• Modern browser with WebAssembly support
• Chrome, Edge, Firefox, or Safari
• Secure context (https:// or localhost)
• Cross-origin isolation headers for threaded ONNX Runtime Web

Voice Cloning

1. Select a language bundle.
2. Choose a built-in voice or upload your own sample.
3. Use a short clean reference clip for best results.
4. Generate directly in the browser.

File Structure

pocket-tts-web/
├── index.html
├── onnx-streaming.js
├── inference-worker.js
├── PCMPlayerWorklet.js
├── EventEmitter.js
├── sentencepiece.js
├── style.css
└── onnx/
    ├── english_2026-04/
    ├── german/
    ├── italian/
    ├── portuguese/
    └── spanish/

License

• Models and bundled voice assets inherit the Pocket TTS licensing terms from kyutai/pocket-tts.
• Code is Apache 2.0.