How butspeechfit/diarizen works
DiariZen provides pretrained speaker diarization inference plus recipes for self-supervised training, multi-channel operation, and structured pruning/distillation. The principal workflow loads a model hub, accepts an audio file, runs diarization, and can write RTTM output. Pyannote supplies the shared audio/model inference and task framework. Training and pruning links are documented at capability level; unsampled implementations are represented by their exact directories, and uncertain runtime calls are omitted.
Main entry points: trainer_dual_opt.py Dual optimizer trainer · inference.py DiariZen pipeline · trainer_single_opt.py Single optimizer trainer · task.py Task framework
Components 28
Diarization inference
- DiariZen pipeline
diarizen/pipelines/inference.py - VBx clustering
diarizen/clustering/VBx.py
Model architecture
- WavLM diarization
diarizen/models/eend/model_wavlm_conformer_mc.py - Prunable SSL model
diarizen/models/module/wav2vec2/model.py - SSL components
diarizen/models/module/wav2vec2/components.py - Conformer blocks
diarizen/models/module/conformer.py - Distillation model
diarizen/models/pruning/model_distill_prune.py
Training and pruning
- Training recipes
recipes/diar_ssl - Multi-channel recipes
recipes/diar_ssl_mc - Single optimizer trainer
diarizen/trainer_single_opt.py - Dual optimizer trainer
diarizen/trainer_dual_opt.py - Distill-prune trainer
diarizen/trainer_distill_prune.py - Pruning recipes
recipes/diar_ssl_pruning - Checkpoint utilities
diarizen/ckpt_utils.py - Noam updater
diarizen/noam_updater.py - Optimization helpers
diarizen/optimization.py - Trainer utilities
diarizen/trainer_utils.py - Experiment logging
diarizen/logger.py
Pyannote framework
- Audio I/O
pyannote-audio/pyannote/audio/core/io.py - Chunk inference
pyannote-audio/pyannote/audio/core/inference.py - Audio model
pyannote-audio/pyannote/audio/core/model.py - Pipeline framework
pyannote-audio/pyannote/audio/core/pipeline.py - Speaker diarization
pyannote-audio/pyannote/audio/pipelines/speaker_diarization.py - Signal utilities
pyannote-audio/pyannote/audio/utils/signal.py - Task framework
pyannote-audio/pyannote/audio/core/task.py - Training losses
pyannote-audio/pyannote/audio/utils/loss.py - Powerset encoding
pyannote-audio/pyannote/audio/utils/powerset.py
Other components
- Audio user
Connections 28
- Audio user → DiariZen pipeline: submits audio
- DiariZen pipeline → Audio I/O: loads audio
diarizen/pipelines/inference.py - DiariZen pipeline → Speaker diarization: runs diarization
diarizen/pipelines/inference.py - DiariZen pipeline → Signal utilities: uses utilities
diarizen/pipelines/inference.py - Chunk inference → Audio model: loads model
pyannote-audio/pyannote/audio/core/inference.py - Pipeline framework → Chunk inference: manages inference
pyannote-audio/pyannote/audio/core/pipeline.py - Speaker diarization → Pipeline framework: uses framework
- WavLM diarization → Prunable SSL model: uses encoder
diarizen/models/eend/model_wavlm_conformer_mc.py - WavLM diarization → Conformer blocks: uses blocks
diarizen/models/eend/model_wavlm_conformer_mc.py - Prunable SSL model → SSL components: builds from
diarizen/models/module/wav2vec2/model.py - Training recipes → Single optimizer trainer: trains with
- Training recipes → Dual optimizer trainer: trains with
- Multi-channel recipes → Dual optimizer trainer: trains with
- Pruning recipes → Distill-prune trainer: trains with
- Distill-prune trainer → Distillation model: optimizes
diarizen/trainer_distill_prune.py - Pruning recipes → Prunable SSL model: prunes model
recipes/diar_ssl_pruning/apply_pruning.py - Single optimizer trainer → Optimization helpers: uses optimizer
diarizen/trainer_single_opt.py - Single optimizer trainer → Trainer utilities: uses utilities
diarizen/trainer_single_opt.py - Single optimizer trainer → Experiment logging: logs progress
diarizen/trainer_single_opt.py - Single optimizer trainer → Noam updater: uses scheduler
diarizen/trainer_single_opt.py - Dual optimizer trainer → Optimization helpers: uses optimizer
diarizen/trainer_dual_opt.py - Dual optimizer trainer → Trainer utilities: uses utilities
diarizen/trainer_dual_opt.py - Dual optimizer trainer → Experiment logging: logs progress
diarizen/trainer_dual_opt.py - Dual optimizer trainer → Noam updater: uses scheduler
diarizen/trainer_dual_opt.py - Task framework → Training losses: uses losses
pyannote-audio/pyannote/audio/core/task.py - Task framework → Powerset encoding: uses encoding
pyannote-audio/pyannote/audio/core/task.py - Audio model → Task framework: configures task
pyannote-audio/pyannote/audio/core/model.py - DiariZen pipeline → Checkpoint utilities: loads weights
diarizen/pipelines/inference.py