Silero VAD for locating speech segments

How it works

The model scores short audio windows for voice activity. Thresholds and segment rules turn those scores into spans for applications such as transcription or recording control.

What you need

Set the sample rate and choose a short-window or longer-recording workflow. Pick padding and minimum-duration rules using examples of what counts as speech for your application.

What to try first

Run a recording with background music, pauses and quiet speech. Sweep a threshold and compare false triggers with missed speech at phrase boundaries.

Test a microphone sample and an already-normalized recording. Differences in gain or sample rate can change which words the detector clips.

Where to look for failures

A speech detector can clip the first or last sounds of an utterance. Use padding or endpoint rules and test for the audio conditions your application will see.

Related Franklin pages

Check whether the task needs speech detection, precise word timing or speaker segments. Save the output of each stage so a timing error can be traced to its origin.

• Add word timing and optional speaker diarization to a transcription workflow. WhisperX for word-aligned transcripts
• Run speaker diarization pipelines to identify who spoke when in a recording. pyannote.audio for separating speakers by time

For an application where the quality of recording and handling audio matters, see Franklin’s Open Mic AI profile. Open Mic AI

Project documentation

• Current project README
• README at the reviewed revision