livekit-wakeword/v0.1.1

Features

Add livekit-wakeword crate with ONNX-based wake word detection

#926 by @pham-tuan-binh

Summary

• New livekit-wakeword crate with a stateless wake word detection pipeline
• Pipeline: raw PCM audio → mel spectrogram → speech embeddings → classifier scores
• Mel spectrogram and embedding models are bundled at compile time via include_bytes!
• Wake word classifier models (e.g. hey_livekit.onnx) are loaded dynamically from disk at runtime
• Supports multiple classifiers simultaneously, each returning a confidence score (0-1)
• Input sample rate resampling via FIR resampler (supports 16–384 kHz, internally resamples to 16 kHz)
• Pure-Rust ONNX backend via ort-tract (falls back to native ONNX Runtime on aarch64-pc-windows-msvc)
• Fix mel spectrogram post-processing normalization (x/10 + 2) to match the openWakeWord pipeline
• Custom WakeWordError enum replacing Box<dyn Error> in the public API

Test plan

• cargo test -p livekit-wakeword — integration tests exercise the full pipeline
• Validates score output is in [0.0, 1.0] range
• Validates too-short audio returns zero scores
• Positive WAV sample ("Hey LiveKit") scores >= 0.5 threshold
• Negative WAV sample (non-wake-word audio) scores < 0.5 threshold
• Rust scores match Python reference implementation (0.9997 positive, 0.0009 negative)