WASM Streaming Speech Recognition

Real-time streaming speech-to-text transcription running entirely in your browser using Rust and WebAssembly (WASM). This demo processes audio entirely offline on your CPU after downloading a ~950MB speech recognition model.

Demo

demo-video.mp4

Try it live: https://huggingface.co/spaces/efficient-nlp/wasm-streaming-speech

Technologies

• Kyutai STT Model - 1B param streaming speech recognition model for English and French. This demo uses a 4-bit quantized version of the model.
• Candle - Hugging Face's ML framework for Rust
• Rayon - CPU parallelization for Rust
• wasm-bindgen-rayon - WASM bindings for Rayon

Related Projects

This is a research/tech demo. For more accurate cloud transcription and real-time LLM grammar correction, check out Voice Writer.

Performance

Performance varies by device.

• On Apple Silicon or other recent CPUs, it typically runs in real time.
• On older devices, it may not keep up (real-time factor < 1).
• Mobile devices are not supported.

Prerequisites

• Rust, Cargo, wasm32-unknown-unknown

  curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
  rustup target add wasm32-unknown-unknown

• wasm-bindgen-cli

  cargo install wasm-bindgen-cli

• wasm-opt (Binaryen) – optional but recommended

  • macOS: brew install binaryen
  • Ubuntu/Debian: sudo apt install binaryen

• Python 3

• curl

Running Locally

1. Clone the repository:

   git clone https://github.com/lucky-bai/wasm-speech-streaming
   cd wasm-speech-streaming

2. Build the Rust/WASM library:

   ./build-lib.sh

3. Open your browser and go to:

   http://localhost:8000

License

MIT License