Precision-3 now available on API and On-Premise - pyannoteAI Speaker Intelligence and Diarization

All changelogs

Precision-3 now available on API and On-Premise

Our latest diarization model, Precision-3, is available now on the API and for on-premises deployments.

Added

Two new input parameters: vadSensitivity and crosstalkSensitivity. They adjust voice activity and overlap detection independently of one another.
Three new output scores: speakerProbability, speechProbability, and crosstalkProbability. Each answers one question: is this speaker active, is speech present, is there crosstalk.
Three on-premise operating modes: speed, balance, and accuracy. Each sets a different trade-off between processing speed and diarization quality.

Changed

Average diarization error rate drops 10.4% against Precision-2, from 16.02 to 14.35 DER across 15 benchmark datasets. Over half of that improvement (53%) comes from more accurate speaker attribution: the model assigns speech to the wrong speaker 18.7% less often.
On-premise, vadSensitivity replaces vad_aggressiveness and inverts the scale: high sensitivity now means more speech detected, where high aggressiveness previously meant the opposite.

Deprecated

The frame-level confidence output is deprecated in favor of speakerProbability, speechProbability, and crosstalkProbability. turnLevelConfidence is unchanged.

Benchmark results:

Pricing is unchanged. Existing integrations keep running Precision-2 until you set model: "precision-3" or wait for the default switch on October 1. Precision-2 is deprecated on October 15.