Skip to content

Version 3.0.0

Compare
Choose a tag to compare
@hbredin hbredin released this 26 Sep 13:00
· 117 commits to develop since this release

TL;DR

Better pretrained pipeline and model

Benchmark (DER %) v2.1 v3.0
AISHELL-4 14.1 12.3
AliMeeting (channel 1) 27.4 24.3
AMI (IHM) 18.9 19.0
AMI (SDM) 27.1 22.2
AVA-AVD - 49.1
DIHARD 3 (full) 26.9 21.7
MSDWild - 24.6
REPERE (phase2) 8.2 7.8
VoxConverse (v0.3) 11.2 11.3

Major breaking changes

  • BREAKING: pipelines now run on CPU by default
    Use pipeline.to(torch.device('cuda')) to use GPU
  • BREAKING: removed SpeakerSegmentation pipeline
    Use SpeakerDiarization pipeline instead
  • BREAKING: removed support for prodi.gy recipes

Full changelog

Features and improvements

  • feat(pipeline): send pipeline to device with pipeline.to(device)
  • feat(pipeline): add return_embeddings option to SpeakerDiarization pipeline
  • feat(pipeline): make segmentation_batch_size and embedding_batch_size mutable in SpeakerDiarization pipeline (they now default to 1)
  • feat(pipeline): add progress hook to pipelines
  • feat(task): add powerset support to SpeakerDiarization task
  • feat(task): add support for multi-task models
  • feat(task): add support for label scope in speaker diarization task
  • feat(task): add support for missing classes in multi-label segmentation task
  • feat(model): add segmentation model based on torchaudio self-supervised representation
  • feat(pipeline): check version compatibility at load time
  • improve(task): load metadata as tensors rather than pyannote.core instances
  • improve(task): improve error message on missing specifications

Breaking changes

  • BREAKING(task): rename Segmentation task to SpeakerDiarization
  • BREAKING(pipeline): pipeline defaults to CPU (use pipeline.to(device))
  • BREAKING(pipeline): remove SpeakerSegmentation pipeline (use SpeakerDiarization pipeline)
  • BREAKING(pipeline): remove segmentation_duration parameter from SpeakerDiarization pipeline (defaults to duration of segmentation model)
  • BREAKING(task): remove support for variable chunk duration for segmentation tasks
  • BREAKING(pipeline): remove support for FINCHClustering and HiddenMarkovModelClustering
  • BREAKING(setup): drop support for Python 3.7
  • BREAKING(io): channels are now 0-indexed (used to be 1-indexed)
  • BREAKING(io): multi-channel audio is no longer downmixed to mono by default.
    You should update how pyannote.audio.core.io.Audio is instantiated:
    • replace Audio() by Audio(mono="downmix");
    • replace Audio(mono=True) by Audio(mono="downmix");
    • replace Audio(mono=False) by Audio().
  • BREAKING(model): get rid of (flaky) Model.introspection
    If, for some weird reason, you wrote some custom code based on that,
    you should instead rely on Model.example_output.
  • BREAKING(interactive): remove support for Prodigy recipes

Fixes and improvements

  • fix(pipeline): fix reproducibility issue with Ampere CUDA devices
  • fix(pipeline): fix support for IOBase audio
  • fix(pipeline): fix corner case with no speaker
  • fix(train): prevent metadata preparation to happen twice
  • fix(task): fix support for "balance" option
  • improve(task): shorten and improve structure of Tensorboard tags

Dependencies update

  • setup: switch to torch 2.0+, torchaudio 2.0+, soundfile 0.12+, lightning 2.0+, torchmetrics 0.11+
  • setup: switch to pyannote.core 5.0+, pyannote.database 5.0+, and pyannote.pipeline 3.0+
  • setup: switch to speechbrain 0.5.14+