Researchers at the Indian Institute of Science (IISc), Bengaluru, have collaborated with the Digital India BHASHINI Division to develop artificial intelligence-based speech technology for Sourashtra, a language spoken predominantly by the Sourashtra community in Tamil Nadu.
The initiative, announced on October 16, 2026, introduces text-to-speech models designed to convert written Sourashtra text into spoken audio. The development forms part of wider efforts to make digital tools more accessible to Indian languages that have comparatively limited digital datasets and speech-technology resources.
Editorial Insight
Key Highlights
Important points readers should notice.
Issue/Event: Development of AI-based text-to-speech models for Sourashtra.
Location: IISc Bengaluru, Karnataka; the language is spoken predominantly by the Sourashtra community in Tamil Nadu.
Authority/Organisation: Digital India BHASHINI Division, Ministry of Electronics and Information Technology, and Indian Institute of Science, Bengaluru.
Action Taken: Developers created text-to-speech models using the VITS architecture and made a speech corpus containing more than 28 hours of recordings available through AIKOSH.
Impact: The resources could support digital reading, accessibility, language learning and further research into speech technology for underrepresented Indian languages.
The models were developed through collaboration between the Digital India BHASHINI Division, under the Ministry of Electronics and Information Technology, and IISc Bengaluru. They use a neural-network architecture known as VITS, or Variational Inference with adversarial learning for end-to-end Text-to-Speech.
Text-to-speech technology allows a computer system to read written content aloud. It can be used in digital reading tools, language-learning applications, accessibility services and systems that help people interact with information through audio instead of relying exclusively on written text.
The newly developed Sourashtra models include male and female voices. They can process text written in Tamil script as well as Sourashtra script, giving developers and users more flexibility when working with the language.
Editorial Analysis
Why This Matters
AI language tools are often developed first for languages with large amounts of digital text and recorded speech. Languages with fewer available resources can face additional barriers because researchers may need to build datasets before developing reliable applications. The Sourashtra project addresses part of that challenge by combining a language-specific speech corpus with models that can convert text into audio. This can provide a starting point for tools that help people listen to written content and enable developers to experiment with further applications. The project also highlights the importance of making datasets available to the wider research community. Accessible language resources can reduce duplicated work and help researchers evaluate alternative models or develop new tools. The longer-term benefits will depend on how well the generated speech performs in everyday use, whether the resources remain accessible and whether additional applications are developed for the community.
Supporting different scripts is important because languages may be written using more than one script, depending on community practice, educational material and the intended audience. Technology that can handle these variations may make it easier to create digital content for people who use different writing conventions.
The models were trained using a dedicated Sourashtra speech corpus developed at IISc Bengaluru's SPIRE Lab. According to the announcement, the dataset contains more than 28 hours of gender-balanced speech recordings from one male and one female voice artist.
The corpus includes composed sentences and sentences drawn from literature. Transcriptions are available in both Sourashtra and Tamil scripts, providing paired text and speech resources for training and evaluating the models.
The availability of such data is an important part of speech technology development. AI speech systems learn patterns from recorded voices and their corresponding text. When a language has limited digital recordings, creating a suitable dataset can be a major step towards building usable speech tools.
However, the size of a dataset alone does not establish how naturally a model speaks or how accurately it handles every word, dialect or speaking situation. Performance depends on the quality and diversity of the recordings, pronunciation coverage, evaluation methods and how well the model generalises to text it has not encountered before.
The speech corpus, identified as the SPICOR TTS 3.0 Sourashtra Corpus, has been made available through BHASHINI's AIKOSH platform. AIKOSH provides datasets and other resources intended to support the development of Indian-language artificial intelligence systems.
Making the dataset accessible to researchers and developers could help support additional work on Sourashtra speech technology. Universities, language researchers and application developers may be able to use the resources as a foundation for further experiments and applications, subject to the platform's applicable access and usage conditions.
Potential uses include applications that read written material aloud, tools supporting language learning, accessibility features for people who find written content difficult to use, and research into speech synthesis for languages with limited digital resources.
For communities whose languages are less represented in mainstream digital platforms, the availability of speech technology may provide new ways to preserve, access and use written material. It can also help researchers explore how language technologies can be developed for linguistic communities that receive less commercial attention than widely used languages.
The initiative does not mean that a complete digital ecosystem for Sourashtra has already been created. Text-to-speech is one component of language technology. Other capabilities, such as speech recognition, translation, keyboard input, searchable digital archives and conversational systems, require their own development and evaluation.
The project nevertheless provides a practical example of how AI research can be applied to a specific language need. Instead of concentrating exclusively on large, widely used languages, developers can use specialised datasets and models to address gaps in digital access.
BHASHINI is working to expand multilingual digital capabilities through technologies including translation, speech recognition, speech synthesis and transliteration. Its wider objective is to make digital information and public-facing services more accessible across India's linguistic diversity.
The next measure of the Sourashtra models' impact will be their usability in real applications, the quality of generated speech and whether researchers and developers build additional tools around the available resources.
