0003: Use an energy VAD instead of a model

Status: Accepted

Context

voicebench needs to find speech in the user and agent tracks. Neural VADs are more robust to noise, but they add a large dependency, need model downloads, vary between versions, and make results harder to reproduce. The tracks that voicebench analyzes are mostly clean: synthetic or recorded user audio, and TTS output from the agent.

Decision

voicebench uses a streaming energy detector with a start threshold, hysteresis, a minimum speech duration, and a hangover period. It backdates segment boundaries to the first and last active frame, so the confirmation delays do not add latency. Every parameter is configurable per track in the scenario.

Consequences

• The core install needs only NumPy, PyYAML, and websockets, and runs anywhere, including the standalone executable.
• Results are deterministic and easy to explain.
• Background noise above the threshold counts as speech. Scenarios with noisy agents need a higher threshold_db.
• A pluggable model-based VAD remains possible later without changing the metrics.