keithito/tacotron

Implements sequence-to-sequence text-to-speech synthesis with attention-based encoder-decoder architecture, converting text directly to mel-spectrogram frames that are then vocoded into audio. Supports training on public datasets (LJ Speech, Blizzard 2012) with configurable hyperparameters, TensorBoard visualization of alignments, and optional phoneme-level control via CMUDict integration. Includes a demo server for inference and pre-trained checkpoints for immediate synthesis without training.

2,988 stars. No commits in the last 6 months.