VoiceGate
A streaming voice-activity detector and mic/file-to-ASR pipeline with hand-written DSP features, adaptive gating, smoothing, a learned logistic-regression path, and Whisper/Saaras adapters.
VoiceGate is a streaming voice-activity detector and ASR front end built from first principles rather than wrapped around an existing VAD package.
Hand-written signal processing
The pipeline computes DSP features over streaming audio frames, applies adaptive gating to account for changing background conditions, and smooths frame-level decisions so brief noise bursts do not fragment speech.
Learned and deterministic paths
Alongside the deterministic detector, a logistic-regression path provides a lightweight learned alternative. Both feed the same downstream interface, making behaviour easy to compare and test.
Pluggable transcription
Mic and file inputs can flow through Whisper or Saaras adapters without coupling the VAD core to either transcription provider. The implementation is covered by 31 passing tests.