NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #3832 most downloaded on PyPI
State-of-the-art speaker diarization toolkit
Latest stable
4.0.7
Last release 4 months ago
30 Jun 2026
Release timing varies
gaps range from 8 days to 12 months
Nearly every release is documented
notes for 20 of 22 stable releases
1 version withdrawn
withdrawn after publishing
6 years old
23 releases · first in 2020
Drag across the ruler to compare, or tap a release to open it.
fix(pipeline): add missing support for subfolder in Pipeline.from_pretrained
subfolder in Pipeline.from_pretrainedNothing published for this version
improve(telemetry): reduce number of sent packages @litdarya
--average-case option to optimize command @antoinelaurentTask.prepare_data to support saving preprocessors that produce int values in metadata @lylyhanYour coding agent can read these notes before it upgrades. Set up the MCP server →
feat(sample): add transcription of sample file
feat(cli): add --revision option to most CLI commands
--revision option to most CLI commandsCalibration.safe_transform method (supports NaNs as well as any shape)Model.from_pretrained to support lightning 2.6+pyannote-database dependency to 6.1+BREAKING(util): make Binarize.__call__ return string tracks (instead of int ) @benniekiss
Binarize.__call__ return string tracks (instead of int) @benniekisstorch, torchcodec, and torchaudio versions to avoid segmentation faultPipeline.cuda() convenience method @tkanarskypreload option to base Pipeline.__call__ to force preloading audio in memory (@antoinelaurent)permutate faster thanks to vectorized cost functionFull Changelog: 4.0.1...4.0.2
feat: allow passing preloaded pipeline config to get_pipeline
get_pipelinepyannoteai-sdk dependency to 0.3.0OpenTelemetry dependenciesBREAKING(cli): remove deprecated pyannote-audio-train CLI
pyannote/speaker-diarization-community-1 pretrained pipeline relies on VBx clustering instead of agglomerative hierarchical clustering (as suggested by BUT Speech@FIT researchers Petr Pálka and Jiangyu Han).
pyannote/speaker-diarization-community-1 pretrained pipeline returns a new exclusive speaker diarization, on top of the regular speaker diarization.
This is a feature which is backported from our latest commercial model that simplifies the reconciliation between fine-grained speaker diarization timestamps and (sometimes not so precise) transcription timestamps.
from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained(
"pyannote/speaker-diarization-community-1", token="huggingface-access-token")
output = pipeline("/path/to/conversation.wav")
print(output.speaker_diarization) # regular speaker diarization
print(output.exclusive_speaker_diarization) # exclusive speaker diarization
Metadata caching and optimized dataloaders make training on large scale datasets much faster.
This led to a 15x speed up on pyannoteAI internal large scale training.
Change one line of code to use pyannoteAI premium models and enjoy more accurate speaker diarization.
from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained(
- "pyannote/speaker-diarization-community-1", token="huggingface-access-token")
+ "pyannote/speaker-diarization-precision-2, token="pyannoteAI-api-key")
diarization = pipeline("/path/to/conversation.wav")
Pipelines can now be stored alongside their internal models in the same repository, streamlining fully offline use.
Accept pyannote/speaker-diarization-community-1 pipeline user agreement
Clone the pipeline repository from Huggingface (if prompted for a password, use a Huggingface access token with correct permissions)
$ git lfs install
$ git clone https://hf.co/pyannote/speaker-diarization-community-1 /path/to/directory/pyannote-speaker-diarization-community-1
Enjoy!
# load pipeline from disk (works without internet connection)
from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained('/path/to/directory/pyannote-speaker-diarization-community-1')
# run the pipeline locally on your computer
diarization = pipeline("audio.wav")
With the optional telemetry feature in pyannote.audio, you can choose to send anonymous usage metrics to help the pyannote team improve the library.
sox and soundfile audio I/O backends (only ffmpeg or in-memory audio is supported)Python < 3.10use_auth_token to token{pipeline_name}@{revision} syntax in Model.from_pretrained(...) and Pipeline.from_pretrained(...) -- use new revision keyword argument insteadOverlappedSpeechDetection task (part of SpeakerDiarization task)OverlappedSpeechDetection and Resegmentation unmaintained pipelines (part of SpeakerDiarization)huggingface_hub caching directory (PYANNOTE_CACHE is no longer used)Inference now only supports already instantiated modelsmultilabel training in SpeakerDiarization taskwarm_up option in SpeakerDiarization taskweigh_by_cardinality option in SpeakerDiarization taskvad_loss option in SpeakerDiarization taskpyannote-audio-train CLItorchaudio to torchcodec for audio I/Ok-means clusteringwav2vec_frozen option to freeze/unfreeze wav2vec in SSeRiouSS architectureSpeakerDiarization taskhidden option to ProgressHookFilterByNumberOfSpeakers protocol files filterCalibration class to calibrate logits/distances into probabilitiesDetectionErrorRate, SegmentationErrorRate, DiarizationPrecision, and DiarizationRecall metricsSSeRiouSS architecture (@clement-pages)SpeakerDiarization training with manual optimization (@clement-pages)uvlightning from pytorch-lightningio.BytesIO bufferdict (@benniekiss)ToTaToNet architecture (@clement-pages)PixIT training with manual optimization (@clement-pages)SpeechSeparation and SpeakerDiarization docstring (@razi-tm).Upcoming major releases of pyannote.{core,database,metrics,pipeline} dependencies will break 3.x branch. Version 3.4.0 pins those dependencies to comp
Upcoming major releases of pyannote.{core,database,metrics,pipeline} dependencies will break 3.x branch.
Version 3.4.0 pins those dependencies to compatible versions.
fix: (really) fix support for numpy==2.x (@metal3d)
setup: drop support for Python 3.8
numpy==2.x (@ibevers)speechbrain==1.x (@Adel-Moumen)pyannote.audio does speech separation: multi-speaker audio in, one audio channel per speaker out!
pyannote.audio does speech separation: multi-speaker audio in, one audio channel per speaker out!
pip install pyannote.audio[separation]==3.3.0
PixIT joint speaker diarization and speech separation task (with @joonaskalda)ToTaToNet joint speaker diarization and speech separation model (with @joonaskalda)SpeechSeparation pipeline (with @joonaskalda)backendsoundfile backendmax_speakers is set to 1feat(task): add option to cache task training metadata to speed up training (with @clement-pages)
receptive_field, num_frames and dimension to models (with @Bilal-Rahou)fbank_only property to WeSpeaker modelsPowerset.permutation_mapping to help with permutation in powerset space (with @FrenchKrab)pyannote.audio.sample.SAMPLE_FILEreduce option to diarization_error_rate metric (with @Bilal-Rahou)Waveform and SampleRate preprocessorstorch.Tensor support in ArtifactHookPowerset docstring (with @lukasstorck)diarization_error_rate metricModel and nn.Module attributes in Pipeline.to(device)torchaudio >= 2.2.0Model.example_output in favor of num_frames method, receptive_field property, and dimension propertypyannote/speaker-diarization-3.1 (by @simonottenhauskenbun)Providing num_speakers to `pyannote/speaker-diarization-3.1` now works as expected.
Providing num_speakers to pyannote/speaker-diarization-3.1 now works as expected.
num_speakers in pyannote/speaker-diarization-3.1 pipeline`pyannote/speaker-diarization-3.1` no longer requires unpopular ONNX runtime
pyannote/speaker-diarization-3.1 no longer requires unpopular ONNX runtime
TimingHook for profiling processing timeArtifactHook for saving internal stepsHooks"soft" option to Powerset.to_multilabelSpeakerDiarizationAgglomerativeClustering to honor num_clusters when providedmax_speakers or detected num_speakers in SpeakerDiarization pipelinefbank on GPU when requestedWeSpeakerPretrainedSpeakerEmbedding to ONNXWeSpeakerPretrainedSpeakerEmbeddingonnxruntime dependency.
You can still use ONNX hbredin/wespeaker-voxceleb-resnet34-LM but you will have to install onnxruntime yourself.logging_hook (use ArtifactHook instead)onset and offset parameter in SpeakerDiarizationMixin.speaker_count
You should now binarize segmentations before passing them to speaker_count`pyannote/speaker-diarization-3.0` is now much faster when sent to GPU.
pyannote/speaker-diarization-3.0 is now much faster when sent to GPU.
import torch
from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.0")
pipeline.to(torch.device("cuda"))
onnxruntime to onnxruntime-gpuMuch better overlapping speech detection with _powerset_ pyannote/segmentation-3.0
| 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 |
pipeline.to(torch.device('cuda')) to use GPUSpeakerSegmentation pipeline
Use SpeakerDiarization pipeline insteadprodi.gy recipespipeline.to(device)return_embeddings option to SpeakerDiarization pipelinesegmentation_batch_size and embedding_batch_size mutable in SpeakerDiarization pipeline (they now default to 1)SpeakerDiarization taskSegmentation task to SpeakerDiarizationpipeline.to(device))SpeakerSegmentation pipeline (use SpeakerDiarization pipeline)segmentation_duration parameter from SpeakerDiarization pipeline (defaults to duration of segmentation model)FINCHClustering and HiddenMarkovModelClusteringpyannote.audio.core.io.Audio is instantiated:
Audio() by Audio(mono="downmix");Audio(mono=True) by Audio(mono="downmix");Audio(mono=False) by Audio().Model.introspection
If, for some weird reason, you wrote some custom code based on that,
you should instead rely on Model.example_output.Version 2.1.x introduces a major overhaul of pyannote.audio default speaker diarization pipeline, made of three main stages:
Version 2.1.x introduces a major overhaul of pyannote.audio default speaker diarization pipeline, made of three main stages:
More details in the attached technical report.
feat: pretrained pipelines (and models) on Huggingface model hub
fix: make sure master branch is used to load pretrained models
Nothing published for this version
last release before complete rewriting
Nothing published for this version
One column per quarter.