Post
2,765
Introducing Precision-3!
Note: We'll walk through the model and take questions in our webinar this Thursday, October 8 at 5 PM CEST : https://app.livestorm.co/pyannote-ai/precision-3
Different tasks want different diarization errors: TTS data prep needs overlap removed, transcription needs every quiet word kept. Precision-3 lets you shift the operating point at inference time, without retraining, via vad_sensitivity and crosstalk_sensitivity (log-prior shifts, not post-hoc thresholds).
Results across the 11 DIHARD domains:
- VAD F1: 95.8% (vs 95.0% community-1, 91.9% Silero VAD)
- Overlap detection F1: 59.7% vs 48.2%
- The optimal VAD setting varies by domain, from −0.8 (Restaurant) to +2.0 (Audiobooks)
- Averaged over 15 datasets, DER drops from 16.02 (Precision-2) to 14.35, mostly from lower speaker confusion.
Write-up: https://pyannote.ai/blog/tune-vad-crosstalk-speaker-diarization
Let us know what you're building with pyannote!
Note: We'll walk through the model and take questions in our webinar this Thursday, October 8 at 5 PM CEST : https://app.livestorm.co/pyannote-ai/precision-3
Different tasks want different diarization errors: TTS data prep needs overlap removed, transcription needs every quiet word kept. Precision-3 lets you shift the operating point at inference time, without retraining, via vad_sensitivity and crosstalk_sensitivity (log-prior shifts, not post-hoc thresholds).
Results across the 11 DIHARD domains:
- VAD F1: 95.8% (vs 95.0% community-1, 91.9% Silero VAD)
- Overlap detection F1: 59.7% vs 48.2%
- The optimal VAD setting varies by domain, from −0.8 (Restaurant) to +2.0 (Audiobooks)
- Averaged over 15 datasets, DER drops from 16.02 (Precision-2) to 14.35, mostly from lower speaker confusion.
Write-up: https://pyannote.ai/blog/tune-vad-crosstalk-speaker-diarization
Let us know what you're building with pyannote!