Add streaming Silero VAD runner for real-time speech detection

[seyeong-han]

Summary

Add a streaming CLI entry point (silero_vad_stream_runner) for the Silero VAD model that enables real-time, frame-by-frame voice activity detection from stdin. This powers the "hey torch" wake-up feature in the Voxtral Realtime macOS app.

Changes

New: silero_vad_stream_runner

A CLI that reads 16kHz mono float32 PCM from stdin and outputs per-frame speech probabilities via a line protocol:

READY
PROB 0.032 0.039
PROB 0.064 0.808
PROB 0.096 0.980

This enables any app to run Silero VAD as a subprocess — pipe audio in, parse probabilities out. The Voxtral macOS app uses this for hands-free wake-up detection.

New: Streaming API on SileroVadRunner

• reset_stream() — re-initialize LSTM state and context buffers
• process_frame(audio_data, num_samples) — process a single 512-sample chunk, return speech probability, carry LSTM state forward

The existing detect() method now uses process_frame() internally, so offline and streaming paths share the same inference code.

Build changes

• CMakeLists.txt — add silero_vad_stream_runner target alongside silero_vad_runner
• Remove unnecessary extension_llm_runner link dependency that caused string_view ambiguity with sentencepiece headers
• Makefile silero-vad-cpu target — build both runners, configure with -DEXECUTORCH_BUILD_EXTENSION_LLM_RUNNER=OFF
• README.md — document streaming usage, architecture, and line protocol

Usage

# Build
make silero-vad-cpu

# Run on a file
ffmpeg -i input.wav -ar 16000 -ac 1 -f f32le pipe:1 | \
  ./cmake-out/examples/models/silero_vad/silero_vad_stream_runner \
    --model_path silero_vad.pte

# Run on live mic
ffmpeg -f avfoundation -i ":0" -ar 16000 -ac 1 -f f32le pipe:1 | \
  ./cmake-out/examples/models/silero_vad/silero_vad_stream_runner \
    --model_path silero_vad.pte

Test plan

• make silero-vad-cpu builds both silero_vad_runner and silero_vad_stream_runner
• Offline runner produces same results as before on test WAV files
• Stream runner on saved audio produces correct speech probabilities (0.8-1.0 on speech, <0.01 on silence)
• Stream runner on live mic input detects speech in real-time
• Integration tested with Voxtral Realtime macOS app wake-up flow

Authored with assistance from Claude.

Made with Cursor

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18507

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.