BHASHINI and IISc Build Sourashtra Text-to-Speech Models

India's language technology push has reached one more small language. The Digital India BHASHINI Division (DIBD), which works under the Digital India Corporation in the Ministry of Electronics and Information Technology (MeitY), has built text-to-speech (TTS) models for Sourashtra together with the Indian Institute of Science (IISc), Bengaluru.

The language

Sourashtra is a minority Indo-Aryan language spoken mainly by the Sourashtra community in Tamil Nadu. The 2011 Census counts about 2.48 lakh speakers. Until now it had very little presence in the digital world and almost no speech technology of its own, which is what makes it a low-resource language in this field.

What was built

Four models have been released in all. They use the VITS architecture, short for Variational Inference with adversarial learning for end-to-end Text-to-Speech, and turn written text straight into spoken Sourashtra. There are two voices, one male and one female, and each voice is offered in two versions, one for text written in the Tamil script and one for text in the Sourashtra script, so a user may work in either script.

How the models were trained

The recordings mix composed sentences with sentences drawn from literature, and the transcriptions exist in both the Sourashtra and the Tamil script. The models read text at the character level, so no separate phonemizer is needed to convert spelling into sounds before speech is produced.

Where the data sits

The corpus has been placed on AIKOSH, the BHASHINI platform that holds language datasets and resources for building Indian language technology. It is listed there as the SPICOR TTS 3.0 Sourashtra Corpus, which means other researchers and developers can use the same data to build further tools.

Why it matters

The work goes beyond preserving a language. Speech-enabled tools can read Sourashtra text aloud, help students and learners of the language, and make digital content usable for people who depend on spoken rather than written material. For researchers it is a base for further work on low-resource speech synthesis.

The wider BHASHINI effort

BHASHINI is building a national ecosystem for multilingual digital inclusion using artificial intelligence. It supports 36 Indian text languages, 23 Indian voice languages and 35 international languages, and offers translation, speech recognition, speech synthesis and transliteration services that government and public-facing digital platforms can plug into.