CIIL launches open-source speech AI ecosystem for Indian languages

The Central Institute of Indian Languages (CIIL), Mysuru, has launched Anuvani, an open-source ecosystem of speech artificial intelligence (AI) models designed to strengthen voice-based technologies across India’s linguistic diversity.

The initiative, announced on Tuesday by the CIIL under the Department of Higher Education, Ministry of Education, Government of India, begins with two state-of-the-art models: Anuvani-LID, which identifies the language being spoken, and Anuvani-Diarization, which identifies speakers and determines when they speak. Both models support all 22 constitutionally scheduled Indian languages and Indian English.

The models are being released with peer-reviewed research papers, open-source model weights, on-device mobile toolkits and production-grade NVIDIA Triton Inference Server kits, enabling developers, researchers, startups and public institutions to use and deploy the technology without proprietary licensing barriers, according to CIIL.

In a press release, the institute said the models address challenges faced by speech AI systems in India, where conversations often involve closely related languages, diverse regional accents, overlapping speakers and noisy acoustic environments.

“For over five decades, CIIL has worked to preserve, document and advance India’s rich linguistic heritage. With the launch of the Anuvani speech AI ecosystem, we are bringing that deep linguistic foundation into the frontier AI era,” said Basavaraja Kodagunti, Director, CIIL.

He said the release of research papers, model weights and deployment toolkits would help ensure that Indian languages remain a shared public resource and enable citizens, developers and institutions to participate in the digital future in their mother tongues.

Tariq Khan, Academic Secretary, CIIL, Mysuru, said Anuvani-LID, built on a 115.6-million-parameter Conformer architecture, was trained on 54,643 hours of Indian speech comprising more than 31 million utterances.

The second model, Anuvani-Diarization, has 76.1 million parameters and is designed to distinguish speakers in phone calls, meetings, recordings made in large rooms and broadcast audio, including situations involving rapid speaker changes and overlapping speech.

“In a zero-shot evaluation on IndicDiarBench, covering 1,164 recordings totalling 107.7 hours across all 22 scheduled languages, the model recorded the lowest Diarization Error Rate among 18 systems tested by CIIL. The evaluation used a strict scoring standard that included overlapping speech,” Dr. Khan said.

The institute said it had engineered the Anuvani releases as complete end-to-end packages so that organisations of any size could deploy Indian-language AI immediately, from cloud clusters to offline smartphones.