BHASHINI IISc Sourashtra VITS TTS Models

This release contains four end-to-end text-to-speech models for Sourashtra (ISO 639-3: saz). Sourashtra is a minority Indo-Aryan language spoken mainly by the Sourashtra community in Tamil Nadu, India. It has little digital presence and few speech technology resources, and these models are meant to help with documenting, preserving and using the language.The models use the VITS architecture (Variational Inference with adversarial learning for end-to-end Text-to-Speech), trained with the Coqui TTS framework. VITS turns text directly into audio in a single model, with no separate vocoder, so it produces natural-sounding speech in one step.There are two voices, one male and one female. Each voice comes in two versions: one that takes Tamil-script text and one that takes Sourashtra-script text. Sourashtra is written in Tamil script in everyday use and also has its own traditional script, so you can use whichever one your text is in. Both versions for a voice produce Sourashtra speech from the same speaker. The models were trained on a custom-annotated Sourashtra speech corpus of paired audio recordings and transcriptions in both scripts. The male voice was trained on about 10,000 utterances and the female voice on about 11,400. The transcriptions were cleaned before training. The cleaning applied Unicode normalization, resolved alternative forms of numbers, removed punctuation, and left out rows containing Latin characters. Each model is a single-speaker VITS model that works directly on characters, so no phonemizer is needed. Audio is generated at 22,050 Hz. Training ran for up to 500 epochs with a batch size of 16 and mixed precision: about 300,000 steps for the male models and 340,000 steps for the female models. The models are intended for reading Sourashtra text aloud, language learning tools, accessibility applications, and research on low-resource speech synthesis. Each model has one voice and a fixed speaking style. Words or characters that didn't appear in the training data may be mispronounced. Numbers, Latin-script text and abbreviations aren't expanded automatically, so write them out as Sourashtra words first.