Gnani AI launches Prisma v3.0 speech model for code-mixed Indian calls

Gnani AI, an Indian voice artificial intelligence (AI) company, launched Prisma v3.0, a speech-to-text model built for Indian phone calls, on Wednesday, October 28, 2026. The company says the model transcribes ten languages, nine Indian languages and English, and copes with speakers who switch between Hindi or another Indian language and English in the middle of a sentence. It is offered to businesses through Gnani's developer interface, and the company claims it makes fewer errors than rival products from ElevenLabs and Microsoft on a public Indian-language test.

Speech recognition, also called automatic speech recognition or ASR, is the technology that turns spoken words into written text. It sits behind voice assistants, automatic captions and, increasingly, the AI "voice agents" that answer customer calls for banks, telecom companies and online shops. India is one of the hardest markets for this technology. People speak dozens of languages and dialects, often mix two or three in a single sentence, and many calls still travel over narrow-band phone lines that carry only part of the sound of a voice.

Most leading speech models were first trained on clean recordings of English and other widely spoken world languages. They tend to make more mistakes on Indian accents, regional languages and code-mixed speech. That gap has created room for Indian companies that train their systems on local data. Gnani's pitch is that its model has heard the kind of audio that Indian businesses actually record: crowded, noisy, switching between languages and passed through call-centre systems.

The main measure in this field is the word error rate. It counts how many words a system gets wrong, leaves out or adds, as a share of the words actually spoken. A lower number is better: a 16.9% rate means roughly one word in six is wrong on that test. These are Gnani's own figures, measured by the company, and have not yet been confirmed by independent testers. Benchmarks also do not always reflect how a system performs on a particular customer's calls.

Gnani is not alone. Another Indian AI lab, Chariot, has also published a speech model for ten Indian languages and Indian English, with its own claims of fewer errors than models from Google and Sarvam. The competing claims show how crowded the Indian-language voice market has become, as start-ups and global companies race to supply the systems that Indian businesses will use to automate millions of customer conversations.

Short-term: Businesses that run large customer-service operations in Indian languages, such as banks, insurers, lenders and telecom companies, get one more option to test for call transcription and voice bots, and can compare it with global and Indian rivals.

Long-term: Better speech recognition in Indian languages could make digital services easier to use for people who are more comfortable speaking than typing, or who do not use English. It could also speed up the automation of call-centre work, which would change the kind of jobs available in that sector.

Who is affected: Indian enterprises that use voice AI in customer service; call-centre workers whose tasks may be automated or changed; developers building apps in Indian languages; and users who want to speak to digital services in their own language.

Questions and Answers