tts-web

Browser-native text-to-speech running 100% client-side via Rust/WASM.

Try the demo →

Disclaimer: Experimental port. Model from Pocket TTS by Kyutai Labs.

Requirements

• Modern browser with WebAssembly support

Quick Start

# 1. Build WASM package
wasm-pack build crates/tts-wasm --target web

# 2. Start dev server
bun web/serve.mjs

Architecture

• TTS model (crates/tts-wasm/): Pocket TTS compiled to WebAssembly via Candle. Generates speech from text tokens using a voice embedding.
• mimi-rs: Shared Rust library for the Mimi audio codec (encoder + decoder + streaming transformer). Used by both tts-web and stt-web.
• Web UI (web/): Web Worker orchestrates model loading and generation, streams audio chunks back to the main thread for real-time playback.

Quantization

The model ships as a GGUF Q8_0 file (~130MB). Weights are loaded directly as Q8_0 via candle's QMatMul, keeping ~97M quantized parameters in memory (~103MB vs ~388MB F32) and reducing memory bandwidth ~4x per inference step.