KinglittleQ/GST-Tacotron

Implements Global Style Tokens (GST) with a Tacotron2 encoder-decoder architecture to learn unsupervised prosodic representations from speech spectrograms, enabling fine-grained style control and transfer across speakers without explicit style labels. Supports multispeaker datasets including Blizzard, with modular components for encoder/decoder networks, loss computation, and mel-spectrogram synthesis. Designed for PyTorch with configurable hyperparameters and dataset preprocessing pipelines for training end-to-end TTS models.

374 stars. No commits in the last 6 months.