Repository navigation
2026.10.3
This release adds configurable speaker embedding models for diarization and improves model downloads, caching and startup.
Improvements
-
Choose from 21 supported models with
WHISPER_DIARIZE_EMBEDDING_MODEL. WeSpeaker ResNet34-LM and NeMo TitaNet large are alternatives for English audio. Compare results on your own recordings and retune the clustering threshold after switching. -
Configure the embedding model and clustering threshold in the Colab notebook.
-
Pre-download another supported embedding without changing the active model:
docker exec whisper whisper_manage --downloaddiarize wespeaker_en_voxceleb_resnet34_LM.onnx -
Docker and Colab startup now allow more time for diarization model downloads and retries. The
whisper_manage --downloadmodelcommand now returns a nonzero exit status when a download fails.
Upgrade notes
The default embedding model, cache filenames and API response formats remain unchanged. Diarization still requires WHISPER_DIARIZATION=true, runs on the CPU and applies only to non-streaming requests. Invalid embedding filenames stop startup before model loading, even when diarization is disabled.
- Cached diarization models are now checked by file size and SHA-256. Valid files, including symlinked models, are reused. Invalid files are preserved and identified in the logs; move the affected cache entries aside and restart with downloads enabled. If you previously swapped weights under another model's filename, restore the official file or rename the replacement to its correct supported filename and select it explicitly. If a diarization model cannot be verified, downloaded or loaded, diarization is disabled while transcription remains available.
- Offline deployments:
WHISPER_LOCAL_ONLYnow also blocks automatic diarization downloads. Pre-cache both the segmentation model and the selected embedding model. Explicitwhisper_manage --downloaddiarizeprefetch remains permitted. - Manual script updates: use
api_server.py,diarizer.pyand the newdiarization_models.pyfrom the same revision. Docker images and the notebook include the helper.
See the diarization guide for model selection and recovery details.
Thanks to @surfnterd for configurable embedding model support in #16.
This release identifies a source and documentation snapshot. Docker images continue to update independently through the moving :latest and :cuda tags.