Kyubyong/tacotron

Implements encoder-decoder architecture with attention mechanism and Griffin-Lim vocoder for mel-spectrogram-to-waveform conversion, trained on multiple public datasets (LJ Speech, audiobooks, Bible recordings). Includes heavily documented training pipeline with bucketed batches, Noam learning rate scheduling, and gradient clipping, plus pre-trained checkpoints and attention visualization tools for monitoring alignment quality during training.

1,833 stars. No commits in the last 6 months.