How multysquid/kitsune-transcribe works
Kitsune-Transcribe builds and evaluates compact Japanese speech-recognition students by preparing audio data, generating teacher labels, selecting examples, pruning a teacher model, and distilling and evaluating the result. Its documented study workflow adds Parakeet-based CTC students and pre-registered comparisons. The graph follows the README’s stated artifact flow and documented script responsibilities; sampled source excerpts are partial, so internal call wiring not established by them is omitted.
Main entry points: store.py Label stores · 01_prepare_data.py Prepare shards · student.py Student checkpoint · 02_teacher_pass.py Teacher labeling
Components 24
Data and labels
- Prepare shards
scripts/01_prepare_data.py - Teacher labeling
scripts/02_teacher_pass.py - Parakeet labeling
scripts/02p_parakeet_pass.py - Second opinion
scripts/02b_second_opinion.py - Select examples
scripts/make_selection.py - Label stores
kitsune/store.py
Student models
- Prune teacher
scripts/03_build_student.py - Build CTC student
scripts/03c_build_ctc_student.py - Student checkpoint
kitsune/student.py - CTC student
kitsune/ctc_student.py
Training and evaluation
- Training data
kitsune/trainset.py - Distill student
scripts/04_distill.py - Audio features
kitsune/features.py - Speech generation
kitsune/generation.py - Evaluate checkpoint
scripts/05_evaluate.py - Evaluation metrics
kitsune/evaluate.py
Study and operations
- Study rules
kitsune/prereg.py - Study statistics
kitsune/study_stats.py - Study report
tools/study_report.py - Label box control
vast/label.py
Other components
- Researcher
- Speech datasets
- Cohere teacher
- Parakeet teacher
Connections 32
- Researcher → Prepare shards: starts data prep
- Speech datasets → Prepare shards: provides audio
- Prepare shards → Teacher labeling: writes shards
- Teacher labeling → Cohere teacher: runs inference
- Teacher labeling → Label stores: writes teacher labels
- Prepare shards → Parakeet labeling: provides shards
- Parakeet labeling → Parakeet teacher: runs inference
- Parakeet labeling → Label stores: writes Parakeet labels
- Prepare shards → Second opinion: provides utterances
- Second opinion → Label stores: writes hypotheses
- Label stores → Select examples: supplies labels
- Select examples → Prune teacher: provides selection
- Cohere teacher → Prune teacher: initializes student
- Prune teacher → Student checkpoint: writes checkpoint
- Parakeet teacher → Build CTC student: provides architecture
- Build CTC student → CTC student: writes checkpoint
- Label stores → Training data: provides targets
- Student checkpoint → Training data: provides model metadata
- Training data → Distill student: provides batches
- Student checkpoint → Distill student: loads checkpoint
- Distill student → Student checkpoint: saves checkpoints
- Label stores → Audio features: provides audio
- Audio features → Speech generation: provides features
- Student checkpoint → Evaluate checkpoint: loads checkpoint
- Label stores → Evaluate checkpoint: provides eval data
- Evaluate checkpoint → Evaluation metrics: uses scoring
- Evaluate checkpoint → Study report: writes eval tables
- Study rules → Study report: defines comparisons
- Study report → Study statistics: uses statistics
- Label box control → Teacher labeling: runs labeling
- Label box control → Parakeet labeling: runs labeling
- Label box control → Second opinion: runs labeling