How oruk-ai/orukeet works
Orukeet is both a multilingual speech-recognition product and a model-development repository. Its user workflow installs a verified model/runtime, loads a recording through the Python API or persistent worker, decodes it with the native recognizer, and returns text (and word timing data). The repository also contains the Gabor fitting and fine-tuning work, model export, and evaluation programs that support the released recognizer. Training and evaluation subworkflows are only partly evidenced by the supplied excerpts and index; their edges are limited accordingly.
Main entry points: model.py Python recognizer · train.py Model fine-tuning · run.py Unseen evaluation · install.py Runtime installer
Components 23
Data and training
- Prepared speech data
training - Dataset audit
training/audit_prepare.py - Unseen coverage
evaluation/unseen/prepare_coverage.py
Model development
- Gabor fitting
training/gabor_half/fit.py - Frozen Gabor kernels
training/gabor_half/frozen.py - Model fine-tuning
training/librispeech_ft/train.py - Accent adaptation
training/regression_ft/train.py - Checkpoint selection
training/selection/seal_confirmation.py
Evaluation
- ASR evaluation
evaluation/standard_asr/run.py - Unseen evaluation
evaluation/unseen/run.py - Scoring metrics
evaluation/unseen/metrics.py
Export and release
- ONNX export
export/onnx/export_orukeet.py - Native export
evaluation/export_oruk_native.py
Local inference
- Runtime installer
src/orukeet/install.py - Audio preparation
src/orukeet/audio.py - Python recognizer
src/orukeet/model.py - Persistent worker
src/orukeet/worker.py - Native recognizer
src/orukeet/nvidia.py
Other components
- Model developer
- Transcription user
- Audio recording
- Transcript
- Hugging Face
Connections 27
- Model developer → Dataset audit: audits data
training/audit_prepare.py - Dataset audit → Prepared speech data: seals data
training/audit_prepare.py - Model developer → Gabor fitting: fits kernels
- Gabor fitting → Frozen Gabor kernels: provides fits
training/gabor_half/frozen.py - Model developer → Model fine-tuning: trains model
- Model fine-tuning → Frozen Gabor kernels: uses kernels
training/librispeech_ft/train.py - Model fine-tuning → Prepared speech data: trains on
training/librispeech_ft/train.py - Model developer → Accent adaptation: adapts model
- Accent adaptation → Frozen Gabor kernels: uses kernels
training/regression_ft/train.py - Model developer → Checkpoint selection: selects checkpoint
- Unseen coverage → Prepared speech data: prepares coverage
- Model developer → ASR evaluation: evaluates models
- Model developer → Unseen evaluation: runs coverage
- Unseen evaluation → Scoring metrics: scores results
- Unseen evaluation → Unseen coverage: reuses sealed set
- Model developer → ONNX export: exports model
- Model developer → Native export: exports model
- Transcription user → Runtime installer: installs runtime
README.md - Hugging Face → Runtime installer: serves weights
README.md - Runtime installer → Transcription user: writes receipt
README.md - Transcription user → Python recognizer: submits audio
README.md - Audio recording → Python recognizer: provides input
README.md - Python recognizer → Audio preparation: prepares audio
- Python recognizer → Persistent worker: uses worker
- Persistent worker → Native recognizer: loads and transcribes
src/orukeet/worker.py - Native recognizer → Transcript: returns recognition
README.md - Python recognizer → Transcript: returns text
README.md