Bengaluru-based Gnani AI launches Prisma v3.0 speech-to-text model for Indian telephone audio

Bengaluru-based frontier AI company Gnani AI has announced the launch of Gnani Prisma v3.0, its latest speech-to-text (STT) model designed specifically for Indian telephone audio and heavily code-mixed speech. The newly released model supports nine Indian languages alongside Indian English from a single framework.

Prisma v3.0 introduces four transcription modes and five controls that dictate how transcripts are written. According to the company, the model builds upon its existing speech models, which currently power 50 million voice interactions daily. It is now accessible to developers via REST, WebSocket streaming, and batch processing through Gnani's developer API.

In evaluations conducted by Gnani AI using the official Vistaar scoring protocol, Prisma v3.0 recorded a macro word error rate (WER) of 16.9 per cent across nine Indian languages. This compares to 19.98 per cent recorded by ElevenLabs Scribe v2 in the same evaluation. Additionally, the company stated that Prisma v3.0 achieved a lower WER on each of the seven individual Vistaar datasets. For audio clips longer than three seconds, the automatic language identification feature correctly identified the spoken language over 98 per cent of the time.

The platform's transcription modes allow different enterprise teams to extract varying formats from the exact same audio file. The 'Verbatim' mode preserves filler words and repetitions for audit and compliance records, while 'Transcribe' generates clean text in each language's native script with numerical figures, dates, and percentages formatted in digits. The 'Code-mix' mode writes every word in the script of the language it was spoken in, retaining English terms in English. Meanwhile, 'Transliterate' renders the entire transcript using English letters while maintaining the original spoken language.

Ganesh Gopalan, Co-founder and CEO of Gnani AI, noted that Indian business communications frequently shift between languages within a single sentence. He stated that the new model equips enterprises to transcribe these interactions accurately with necessary names and numbers.

Bharath Shankar, Chief Product and Engineering Officer at Gnani AI, added that code-mixed speech serves as the default communication format in India. He highlighted that developers can pass up to 200 keywords per request to ensure product names are transcribed correctly, alongside handling spoken times and native terminology.

"The launch of domain-specific speech models tailored for regional and code-mixed languages highlights an important shift in enterprise technology for the Indian market. Voice interactions in India routinely involve a mix of multiple languages and local terminology, which standard global models often struggle to process accurately. By offering customizable transcription modes and developer APIs, companies like Gnani AI are addressing practical compliance and operational needs for businesses that rely heavily on telephone customer interactions." — Dr. Shishir Gupta, Founder & CEO, StartupLanes