How b04901014/uuvc works
UUVC is a research system for one-shot speech prosody and speaker conversion. Users provide source and target speech; inference extracts speech units and speaker information, predicts timing and prosodic features, generates a mel spectrogram, and vocodes it to audio. The repository also supports feature preprocessing, model training, and a separate HiFi-GAN vocoder-training workflow. The README documents these capabilities; connections for unsampled modules are included only where the README or SOURCE INDEX supports them.
Main entry points: trainer.py Conversion trainer · inference.py Conversion inference · s2rv.py Speech representation · u2mel.py Unit-to-mel model
Components 24
Feature preparation
- Feature extraction
s2u.py - VCTK splits
make_data_vctk.py
Speech conversion
- Conversion inference
inference.py - Exact-pitch inference
inference_exact_pitch.py - HuBERT unit encoder
- Speech representation
models/s2rv.py - Transformer layers
models/transformers.py - Prosody predictors
models/vp.py - Unit-to-mel model
models/u2mel.py - Model utilities
models/utils.py
Model training
- Speech dataset
data/dataset.py - Bucket sampler
data/sampler.py - Training entry
train.py - Conversion trainer
trainer.py - Training losses
loss.py
Vocoder
- Vocoder training
vocoder/train.py - Vocoder mel data
vocoder/meldataset.py - HiFi-GAN models
vocoder/models.py - Vocoder runtime
vocoder/vocoder.py
Other components
- Researcher
- Speech inputs
- Prepared features
- Generated mel
- Converted audio
Connections 31
- Researcher → Speech inputs: provides
README.md - Speech inputs → Feature extraction: extracts features
README.md - Feature extraction → Prepared features: writes
README.md - Prepared features → VCTK splits: splits
README.md - Prepared features → Speech dataset: loads
- Researcher → Conversion inference: requests conversion
README.md - Speech inputs → Conversion inference: reads pairs
README.md - Conversion inference → HuBERT unit encoder: encodes units
inference.py - Conversion inference → Speech representation: runs encoder
inference.py - Conversion inference → Prosody predictors: predicts prosody
inference.py - Conversion inference → Unit-to-mel model: predicts mel
inference.py - Conversion inference → Generated mel: produces
inference.py - Exact-pitch inference → Speech representation: runs encoder
inference_exact_pitch.py - Exact-pitch inference → Unit-to-mel model: predicts mel
inference_exact_pitch.py - Exact-pitch inference → Generated mel: produces
inference_exact_pitch.py - Speech representation → Transformer layers: uses layers
models/s2rv.py - Prosody predictors → Model utilities: uses blocks
models/vp.py - Unit-to-mel model → Model utilities: uses blocks
models/u2mel.py - Generated mel → Vocoder runtime: vocodes
- Vocoder runtime → HiFi-GAN models: runs generator
vocoder/vocoder.py - Vocoder runtime → Converted audio: returns audio
vocoder/vocoder.py - Researcher → Training entry: starts training
README.md - Training entry → Conversion trainer: invokes
train.py - Conversion trainer → Speech dataset: loads batches
trainer.py - Conversion trainer → Bucket sampler: samples batches
trainer.py - Conversion trainer → Speech representation: trains encoder
trainer.py - Conversion trainer → Prosody predictors: trains predictors
trainer.py - Conversion trainer → Unit-to-mel model: trains mel model
trainer.py - Conversion trainer → Training losses: computes losses
trainer.py - Vocoder training → Vocoder mel data: loads data
vocoder/train.py - Vocoder training → HiFi-GAN models: trains models
vocoder/train.py