voice-activity-detection

Native Rust desktop app that listens to a microphone and shows, in real time, whether somebody is speaking. egui draws the window, nalgebra computes the speech features and Burn runs a small feedforward model on the CPU.

Run

cargo run --release

Or with just: just run (just r), just build (just b), just lists all recipes.

Pick a microphone and click START. The button turns blue while listening and green when speech is detected. Input level and speech confidence are shown below; raise the threshold slider if background noise triggers the indicator.

cargo run --release -- --list-microphones   # show capture devices
cargo run --release -- --model other.json   # use another exported model
cargo run --release -- --preview            # 4 s UI preview, microphone stays off

Requirements

• Rust 1.95 or newer
• Linux desktop session (X11 or Wayland) with OpenGL
• arecord from alsa-utils for audio capture (16 kHz mono PCM, processed in memory, nothing is written to disk)

How it works

Audio is cut into frames and turned into 62 features: windowed DFT, mel filter bank, DCT cepstrum, spectral statistics and polynomial derivatives over nine frames (about 50 ms lookahead). The model in model/speech_detector.json is a feedforward network with BatchNorm and ReLU layers exported from PyTorch; feature normalisation is part of the export. Probabilities are smoothed and a 300 ms hangover keeps the indicator stable between syllables.

The bundled model was trained on a small augmented dataset with automatically generated GMM VAD labels (87.7 % held-out frame accuracy). It is a demonstration model, not a production detector.

Headless inference

The infer binary reads newline-delimited JSON on stdin and answers with per-frame probabilities, useful for checks without a display or microphone:

cargo build --release
python3 -c 'import json; print(json.dumps({"samples": [0.0] * 16000}))' | \
  target/release/infer model/speech_detector.json

Send {"samples": [...]} with at least 1440 samples of 16 kHz mono audio, or {"features": [[...62 numbers...]]} to skip preprocessing.

Tests

cargo test --release

Includes a headless layout check of the window and streaming tests that verify chunk-size independence of the feature pipeline.

License

MIT, see LICENSE.