How kyubyong/dc_tts works
This repository implements DC-TTS, a TensorFlow text-to-speech system. The main workflow prepares speech and transcript data, trains Text2Mel and SSRN as separate stages, then synthesizes new text into audio using the trained models. The graph also shows spectrogram processing and shared neural-network primitives. Relationships are grounded in the supplied source excerpts or README; unsampled implementation details are omitted.
Main entry points: train.py Training graph · synthesize.py Synthesis runner · prepo.py Spectrogram preprocessing · data_load.py Text and batches
Components 13
Data preparation
- Text and batches
data_load.py - Spectrogram preprocessing
prepo.py
Model architecture
- Text2Mel
networks.py - SSRN
networks.py - Neural primitives
modules.py
Training and synthesis
- Model parameters
hyperparams.py - Training graph
train.py - Synthesis runner
synthesize.py
Audio processing
- Spectrogram features
utils.py - Waveform reconstruction
utils.py
Other components
- TTS user
- Speech corpus
- Test sentences
Connections 17
- TTS user → Spectrogram preprocessing: prepares data
README.md - Speech corpus → Spectrogram preprocessing: provides audio
- Spectrogram preprocessing → Text and batches: loads transcripts
prepo.py - Spectrogram preprocessing → Spectrogram features: extracts features
prepo.py - TTS user → Training graph: starts training
README.md - Training graph → Text and batches: loads batches
train.py - Training graph → Text2Mel: builds stage one
train.py - Training graph → SSRN: builds stage two
train.py - Text2Mel → Neural primitives: uses layers
networks.py - SSRN → Neural primitives: uses layers
networks.py - Training graph → Model parameters: reads settings
train.py - TTS user → Synthesis runner: requests synthesis
README.md - Synthesis runner → Text and batches: loads test text
synthesize.py - Test sentences → Synthesis runner: supplies sentences
README.md - Synthesis runner → Training graph: builds inference graph
synthesize.py - Synthesis runner → Waveform reconstruction: reconstructs audio
synthesize.py - Spectrogram features → Waveform reconstruction: supports inversion
utils.py