Bengaluru-based Gnani AI, which builds proprietary foundational models and enterprise agentic AI, has launched Gnani Prisma v3.0, the newest version of its speech-to-text (STT) engine. Announced on October 28, the model has been trained specifically on Indian telephone audio and heavily code-mixed speech, the kind of conversation where a caller can switch from Hindi to English to Marathi inside a single sentence.

A single Prisma v3.0 model handles nine Indian languages, namely Hindi, Bengali, Marathi, Gujarati, Punjabi, Kannada, Telugu, Tamil and Malayalam, along with Indian English. It accepts native 8 kHz telephony audio as well as 16 kHz app and VoIP audio, and succeeds the company’s earlier Prisma 2.5 release.

Benchmarked against a global rival

In Gnani AI’s own evaluation using the official Vistaar scoring protocol, Prisma v3.0 posted a macro word error rate (WER) of 16.9 per cent across the nine Indian languages, compared with 19.98 per cent for ElevenLabs Scribe v2. The company said its model recorded the lower error rate on each of the seven Vistaar datasets. On clips longer than three seconds, its automatic language identification picked the correct spoken language more than 98 per cent of the time.

One call, four kinds of transcript

The headline change in v3.0 is control over how a transcript is written. Teams can choose one of four modes on each request, so a compliance function and a voice agent can read the same call in different forms:

• Verbatim retains every filler word and repetition for audit and compliance records.
• Transcribe produces clean, readable text in each language’s native script, with amounts, dates and percentages rendered as digits.
• Code-mix writes each word in the script of the language it was spoken in, while keeping English terms in English.
• Transliterate renders the whole transcript in English letters while preserving the speaker’s own words.

Five additional controls let developers fine-tune output further, including the ability to pass up to 200 keywords per request so that product and brand names are returned correctly.

“Indian business today runs on phone calls, and those calls move between languages within a single sentence,” said Ganesh Gopalan, Co-founder and CEO, Gnani AI. “Prisma v3.0 gives enterprises the ability to transcribe them the way they were spoken, with the numbers and names a business needs to act on.”

Bharath Shankar, Chief Product and Engineering Officer, Gnani AI, said code-mixed speech is the default in India, and that the new release lets teams decide how each transcript is written, from a verbatim record to code-mixed text in which each language keeps its own script. He pointed out that a spoken Marathi time such as “saade-daha” (half past ten) is now returned as 10:30.

Built on scale

Prisma v3.0 builds on the speech models that already power around 50 million voice interactions a day for Gnani AI’s customers, many of them in banking, financial services, insurance, automotive and telecom. The model is available immediately through Gnani’s developer API over REST, WebSocket streaming and batch processing.

Why it matters

For Indian enterprises running large contact centres, collections desks and voice-agent deployments, generic speech engines trained largely on Western, studio-quality audio have long struggled with noisy phone lines, regional accents and mid-sentence language switching. A home-grown model tuned to that reality, with transparent benchmarks and granular output control, strengthens the case for sovereign, India-first AI infrastructure in the voice layer, an area that underpins customer service automation, regulatory audit trails and the next wave of agentic AI applications.

Be a part of Elets Collaborative Initiatives. Join Us for Upcoming Events and explore business opportunities. Like us on Facebook , connect with us on LinkedIn and follow us on Twitter.

"Exciting news! Elets technomedia is now on WhatsApp Channels Subscribe today by clicking the link and stay updated with the latest insights!" Click here!