About this event
Precision-3: what changed, how we measured it, and how to migrate
If you run diarization in production, this session covers what you need to know before upgrading to Precision-3.
Hervé Bredin, co-founder and CSO of pyannoteAI and creator of pyannote.audio, and Jyoti Bisht, Developer Relations Engineer, will run Precision-3 live, go through the benchmark results dataset by dataset, and answer your questions.
Precision-3 offer a diarization error of 14.35, reduces diarization error by 10.4% compared to Precision-2. More than half of that gain comes from fewer speaker confusion errors, which means fewer segments attributed to the wrong person. It also adds input parameters to control how the model handles speech detection and overlapping speech, and output probabilities you can use to filter or review segments.
What you'll leave with:
- How Precision-3 reduces diarization error by 10.4% compared to Precision-2, and how we ran the benchmarks
- How vadSensitivity changes the output, shown on the same file at three settings, with recommended starting values for broadcast, clinical, call-center, court, and restaurant audio
- How crosstalkSensitivity lets you choose between catching every talk-over and keeping only clean single-speaker regions, and where it stops being safe to raise it
- What the new outputs expose and what you can build on them
- How to get a first result in five lines of code, and how to migrate from Precision-2
- Answers to your questions in a live Q&A
🗓️ Date: October 8, 2026, 5 PM CEST (11 AM ET, 8 AM PT)
📹 Format: Live demo, benchmark walkthrough, Q&A
🗣️The session will lasts about 35 minutes. Register even if you can't attend live. We'll send the recording.
pyannote.ai
Identify who speaks when with pyannote.aiSimply detect, segment, label, and separate speakers in any language.
When your audio sources are complex, contaminated by background noise, affected by external elements, and the performances of your solution depend on them, pyannote brings accuracy and precision.
