slp-rl/HebTTS

Built on a language modeling architecture that operates over discrete speech tokens conditioned on word-piece tokenization, enabling pronunciation inference from undiacriticized Hebrew text through contextual understanding. The system uses an autoregressive (AR) model paired with a non-autoregressive (NAR) decoder, trained on weakly supervised data from HebDB, and supports multi-speaker synthesis with optional Multi Band Diffusion vocoding for enhanced audio quality. Inference is accessible via command-line with flexible input handling (text strings, files, or CSV batch processing) and pre-trained checkpoints are provided.

108 stars. No commits in the last 6 months.