bshall/Tacotron

Implements location-relative attention with dynamic convolution to improve alignment robustness in text-to-mel-spectrogram synthesis, enabling stable training on single GPUs with mixed precision. Integrates with the UniversalVocoder for end-to-end audio generation from text via CMUDict phoneme conversion. Provides pretrained LJSpeech weights and preprocessing utilities for dataset training, with architectural optimizations including gradient clipping and modified learning schedules for efficient single-GPU convergence.

115 stars and 32 monthly downloads. No commits in the last 6 months. Available on PyPI.