Skip to content

Version 4.1.0

Choose a tag to compare

@hbredin hbredin released this 17 Nov 11:53
· 1 commit to main since this release
ff2a176

Version 4.1.0 (2026-11-17)

Breaking changes

  • BREAKING: require Python 3.11 or newer; Python 3.10 is no longer supported
  • BREAKING(io): rename Audio.downmix_and_resample to Audio.select_channels_and_resample
  • BREAKING(io): remove Audio(mono=...); use the file's channel policy instead.
  • BREAKING(cli): remove gap-filling optimization (--optimize) and OptimizedMinDurationOff outputs from the benchmark command

Features and improvements

  • feat(task): support channel: {int | downmix | random | all} for audio loading (requires pyannote.database >= 6.3.0)
  • feat(cli): add support for multi-objective pipeline optimization
  • feat(cli): add --subfolder option to benchmark command @antoinelaurent
  • improve(model): compute WeSpeaker fbank features on MPS instead of falling back to CPU @Isalia20

Fixes

  • fix(io): fix Audio.crop raising when an in-memory waveform is cropped up to its very end @MohammadHijjawi97
  • fix(task): avoid training crashes with spawn/forkserver workers, including
    Python 3.14's default on Linux, by warning and using num_workers=0
  • fix(task): fix balance option crashing training with KeyError: 'metadata' @MohammadHijjawi97
  • fix(model): restore fbank_only mode in WeSpeaker embedding models
  • fix(model): honor the configured linear hidden size in ToTaToNet @antonsoo
  • fix(pipeline): fix VBxClustering crashing when more speakers are requested than there are embeddings to cluster @MohammadHijjawi97
  • fix(pipeline): omit unset token and cache_dir when expanding $model/ references so loaders can inherit runtime options
  • fix(utils): fix SpeakerDiarizationMixin.optimal_mapping reference is a mapping @Kayvan-Zahiri
  • fix(utils): fix discrete_diarization_error_rate returning 0.0 on long recordings (float16 frame counts overflowed) @Kayvan-Zahiri