Gnani AI launches Prisma v3.0 speech-to-text model for Indian languages
Gnani AI, a Bengaluru-based AI company has launched Prisma v3.0, a speech-to-text model trained on Indian telephone audio and code-mixed speech. The model supports nine Indian languages and Indian English with four transcription modes and controls for formatting transcripts.
In an evaluation using the official Vistaar scoring protocol, Prisma v3.0 recorded a macro word error rate (WER) of 16.9% across nine Indian languages, compared with 19.98% for ElevenLabs Scribe v2. Gnani AI said, Prisma v3.0 recorded a lower WER across all seven Vistaar datasets. For audio clips longer than three seconds, its automatic language identification feature correctly identified the spoken language more than 98% of the time.
The model builds on Gnani AI’s existing speech technology, which the company says powers 50 million voice interactions daily. Prisma v3.0 is available through the company’s developer API with support for REST, WebSocket streaming and batch processing.
The model offers four transcription modes. Verbatim retains filler words and repetitions for audit and compliance records, while Transcribe produces readable text in the language’s native script, with numbers, dates and percentages rendered as digits. Code-mix preserves each word in the script of the language spoken, while Transliterate renders the transcript in English letters without translating the words.
In Transcribe and Code-mix modes, inverse text normalisation converts spoken numbers into figures using Indian digit grouping. For example, “three lakh fifty thousand” becomes 3,50,000.
The model is designed for Indian business calls, where speakers frequently switch between languages within a sentence and audio is often transmitted over 8 kHz telephone lines. Such conditions can make transcription difficult for systems that process customer interactions, particularly when names, amounts and other details need to be recorded accurately.
For example, Prisma v3.0 can convert spoken expressions such as the Marathi “सव्वादोन कोटी” and Kannada “ಎರಡೂವರೆ ಕೆಜಿ” into “₹2.25 crore” and “2.5 kg” in Transcribe and Code-mix modes. Verbatim mode retains the original spoken wording.
“Indian business today runs on phone calls, and those calls move between languages within a single sentence,” said Ganesh Gopalan, Co-founder and CEO of Gnani AI. He said Prisma v3.0 allows enterprises to transcribe calls according to their requirements while preserving important details such as numbers and names.
Bharath Shankar, Chief Product and Engineering Officer at Gnani AI, said the model allows teams to choose how transcripts are written, from verbatim records to code-mixed text that preserves each language’s script. Developers can also pass up to 200 keywords per request to improve the transcription of product names and other specialised terms.