STT Bench

The same audio every time, with a known correct transcript for each clip, so the two services are scored on identical input. Diarization is forced off here on both sides — the references contain no speaker labels, so this measures transcription accuracy only.

Clips are replayed through both WebSocket APIs at true speaking speed, exactly as a live call would arrive. A run therefore takes as long as the audio itself. Both services are fed the same stream simultaneously.

Normalization is applied identically to both providers. Toggling it re-scores every stored result.