Gnani AI releases speech recognition built for Indian calls that switch languages

Gnani AI has released Prisma v3.0, a speech-to-text model trained on Indian telephone audio and conversations that switch languages mid-sentence. Available through its developer API, it supports nine Indian languages and Indian English. The release gives businesses four ways to write the same call, from an exact record to cleaned-up text for software to process.

An exact record or a readable transcript

In its October 28 launch post, Gnani describes a problem beyond recognizing words: different systems need different versions of a conversation. A compliance reviewer may need every hesitation preserved, while a voice agent needs readable text. Developers choose the output mode on each request, rather than accepting one fixed transcript style.

• Verbatim preserves fillers, repetitions and false starts. Numbers remain written as spoken words, making this the company’s option for reviewing exactly what someone said.
• Transcribe removes fillers and repetitions, uses native scripts, and turns spoken amounts, dates and percentages into digits and symbols.
• Code-mix writes each language in its own script and keeps English terms in English. Numbers stay in spoken-word form.
• Transliterate writes the words in English letters without translating them into English. The speaker’s language and spoken number forms remain intact.

The distinction is important for amounts. Gnani’s technical post specifies that only Transcribe applies inverse text normalization: converting spoken numbers into conventional written figures. Its example turns “three lakh fifty thousand” into 3,50,000, using Indian digit grouping. The other three modes preserve the spoken wording rather than making that conversion.

A benchmark lead, with uneven language results

Gnani’s competitive case rests on its own evaluation using the official Vistaar scoring protocol, covering nine Indian languages and seven datasets. Word error rate counts substituted, missing and extra words against a reference transcript; lower is better. The company reports a macro average of 16.93% for Prisma, ahead of ElevenLabs Scribe v2, Microsoft Azure STT and Deepgram Nova 3.

16.93%Prisma v3.0

Gnani reports this macro word error rate across seven Vistaar datasets covering nine Indian languages.

19.98%ElevenLabs Scribe v2

Scribe v2’s result comes from the same company-run comparison, not an independent evaluation.

20.51%Microsoft Azure STT

Gnani reports Azure’s average at 20.51%; lower scores indicate fewer word errors.

Deepgram Nova 3 recorded 26.57% in that comparison. Prisma had fewer errors than Scribe v2 and Nova 3 on all seven datasets, but Azure beat it on MUCS. The dataset results therefore support a stronger average score, not a clean sweep against every rival.

Language-level scores show another distinction. Prisma’s reported word error rate was 11.72% in Hindi, 23.54% in Tamil and 27.91% in Malayalam. It beat Scribe v2 and Azure in all nine Indian languages, but its own error rate still differed substantially depending on the language being transcribed.

Names, short replies and telephone noise

Developers can supply up to 200 terms per request to steer recognition toward relevant names and vocabulary. Gnani says this raised recall of supplied terms from 70.9% to 90.3% in its tests. There is a useful limit beside that claim: the evaluation used public read speech with the correct terms supplied, which Gnani calls an upper bound.

Language selection has separate controls. Applications can leave detection automatic, provide two to four candidate languages, or supply an expected-language hint without forcing the output. Gnani reports 98% automatic language-identification accuracy on Vistaar and Svarah clips longer than three seconds. That result is specific to those longer clips, rather than a stated accuracy rate for every brief reply.

The model accepts native 8 kHz telephone audio and 16 kHz app or internet-call audio. Gnani also reports that false text generated from noise-only audio fell from 13.4% with the previous model to 4.2%. That measures a different failure from misspelling a word: producing text when no speech was present.

For deployment, Prisma v3.0 is available through REST requests, WebSocket streaming and batch processing. All four output modes work across those endpoints, so the transcript-style choice is not limited to either live calls or completed recordings.