ivanvovk/durian-pytorch

Encoder-decoder architecture that replaces learned attention with explicit phoneme duration information, eliminating attention instability but requiring duration-aligned datasets. Includes a dual-module design with a backbone synthesizer and simultaneous duration predictor trained with configurable BCE/MSE losses. Provides pretrained models on LJSpeech and integrates with Montreal Forced Aligner for dataset phoneme alignment and duration extraction.

184 stars. No commits in the last 6 months.