Turn-taking for real speech
Let people finish.
Speech has pauses, repetitions, and restarts. StutterTurn is an engineering prototype for comparing when a voice agent decides to answer — examining the space between a pause and a finished turn.
A pause can still be part of a turn.
Inspect what
we have.
The replay-v2 handoff checks the recording-to-event path. It is an engineering check, with human review and model evaluation still ahead.
Recorded-source replay v2
2026-10-10 · Replay v2 supersedes the first replaySaved logs executed on bcf50d1
- Sources
- 5 SEP-28k recordings
- Runs
- 10 replay logs across stock and patient
- Setting changed
- Silence gate: 0.2 s → 1.0 s
- Observed
- Stock decided “complete” inside the labelled gap on 4 of 5 clips; 3 of those replies were cancelled before any audio
- Candidate model
- None exists yet; the console lists one only when the backend supplies it
Labels have not been human-reviewed. These runs do not establish improved interruption rates or response delay. Audio receipt is logged; speaker playback is unverified.
Inspect the pinned replay evidenceSame recording. A different wait.
The console puts the caller and the agent on one timeline, so a turn decision and reported response audio can be inspected separately.
- 01
Choose a recorded source
Replay a SEP-28k clip — a real recording of a person who stutters, identified on screen. Keep that source fixed when you change configurations.
- 02
Compare stock and patient
Start a fresh run for each. Both use the same model; the patient setting waits longer through silence.
- 03
Follow the event record
Inspect speech, holding, turn decisions, and reported audio receipt together. A decision alone does not mean the caller heard a reply.
Watch the turn.
Inspect the record.
The judge view brings the comparison, console, and evidence into one place.
Open judge view