The Digital India BHASHINI Division (DIBD) under Digital India Corporation, Ministry of Electronics and Information Technology (MeitY), in collaboration with the Indian Institute of Science (IISc) Bengaluru, has developed Text-to-Speech (TTS) models for Sourashtra, a minority Indo-Aryan language spoken by approximately 2.48 lakh people according to the 2011 Census. The initiative aims to expand speech technology availability for languages with limited digital resources and support their use in digital applications, language learning, accessibility solutions, documentation, and research.
The newly developed models use the VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech) architecture and are capable of generating Sourashtra speech directly from text. The models include male and female voices, with each voice available for text written in both Tamil script and Sourashtra script. The male voice model was trained on approximately 10,000 utterances, while the female voice model was trained on approximately 11,400 utterances, using a dedicated Sourashtra speech corpus containing more than 28 hours of gender-balanced speech recordings from one male and one female voice artist.
The speech corpus, now available on BHASHINI's AIKOSH platform as the SPICOR TTS 3.0 Sourashtra Corpus, includes composed sentences as well as sentences drawn from literature, with transcriptions available in both Sourashtra and Tamil scripts. The models generate audio at 22,050 Hz and operate directly on character-level text, eliminating the need for a separate phonemizer. This technology can support applications including reading Sourashtra text aloud, language learning, accessibility solutions, and research in low-resource speech synthesis, while contributing to the broader development of digital resources for Sourashtra and other underrepresented Indian languages.