Tomiinek/Multilingual_Text_to_Speech

Built on Tacotron 2, the implementation compares three encoder parameter-sharing strategies: full sharing with adversarial speaker classification, separate per-language encoders, and a hybrid approach using a parameter generator to produce language-specific weights. The model integrates domain adversarial training principles and supports synthesis via WaveRNN vocoding, with pre-trained checkpoints available for code-switching and multilingual scenarios trained on CSS10 and Common Voice datasets.

844 stars. No commits in the last 6 months.