Editorial illustration for Gnani AI's Prisma v3.0 Speech Model Targets Transcription Errors in 10 Languages
Gnani AI's Prisma v3.0 Fixes Multilingual Speech Recognition
Gnani AI's Prisma v3.0 Speech Model Targets Transcription Errors in 10 Languages
Gnani AI rolled out Prisma v3.0 this week, an automatic speech recognition model built to handle a problem most voice tech still trips over: Indians rarely speak in one language for an entire sentence. The Bengaluru startup's new model covers Hindi, Bengali, Marathi, Gujarati, Punjabi, Kannada, Telugu, Tamil, Malayalam and English, and it's tuned specifically for conversations that drift between them mid-sentence.
On the Vistaar benchmark, Prisma v3.0 posted a macro Word Error Rate of 16.93, ahead of ElevenLabs Scribe v2 at 19.98, Microsoft Azure Speech-to-Text at 20.51 and Deepgram Nova 3 at 26.57. Gnani says the model also beats its own predecessor, Prisma 2.5, and was designed to clean up the kind of errors that trip up enterprise voice systems: spoken numbers, brand names, short one-word replies, and language-switching within a single utterance.
The release comes with four distinct transcription modes, Verbatim, Transcribe, Code-mix and Transliterate, each aimed at a different use case for developers building voice applications on top of it. Gnani frames the launch around a distinction it says the industry has been slow to address.
Bengaluru-based Gnani AI has launched Prisma v3.0, an automatic speech recognition (ASR) model designed to transcribe multilingual conversations, including speech that switches between Indian languages and English within a sentence.
Why this matters
For teams building voice products in India, the gap between lab benchmarks and real call-center audio has always been where ASR models fail, mangled brand names, misheard numbers, a caller switching from Kannada to English mid-sentence. Gnani AI's Vistaar numbers, 16.93 WER against ElevenLabs Scribe v2's 19.98, suggest Prisma v3.0 is tuned specifically for that mess rather than for clean, single-language dictation. The four transcription modes matter more than the headline score: Verbatim and Transcribe cover different enterprise needs, while Code-mix and Transliter tackle the code-switching problem that most Western-built models still treat as noise.
We'd want to see independent testing on actual call-center data before taking the benchmark gains at face value, since vendor-reported WER comparisons tend to favor the vendor's own test conditions. Still, a 10-language model built around mixed Indian-English speech, rather than retrofitted from an English-first system, is the right design starting point for this market. Worth watching whether Gnani publishes error breakdowns by language pair, since aggregate WER can hide which languages are actually underperforming.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives AI Agent Wabi Pivots to Messenger, Valued at USD 1.25B
From Archives Big Tech pushes back as AI tracking links agencies to surveillance tools
Common Questions Answered
What languages does Gnani AI's Prisma v3.0 speech recognition model support?
Prisma v3.0 covers ten languages including Hindi, Bengali, Marathi, Gujarati, Punjabi, Kannada, Telugu, Tamil, Malayalam, and English. The model is specifically designed to handle conversations that switch between these Indian languages and English mid-sentence, addressing a common challenge in Indian voice technology.
How does Prisma v3.0 perform compared to other ASR models on the Vistaar benchmark?
Prisma v3.0 achieved a macro Word Error Rate of 16.93 on the Vistaar benchmark, which outperforms ElevenLabs Scribe v2's 19.98 WER. This performance difference demonstrates that Prisma v3.0 is specifically tuned for real-world multilingual conversations rather than clean, single-language dictation.
What specific problem does Prisma v3.0 address for voice technology in India?
Prisma v3.0 tackles the challenge that Indians rarely speak in one language for an entire sentence, often switching between languages mid-conversation. This multilingual code-switching has traditionally been a major source of transcription errors in call-center audio and other real-world voice applications.
Why is Prisma v3.0's performance more relevant for Indian call-center applications than traditional ASR benchmarks?
Traditional ASR models often perform well in lab conditions with clean, single-language audio but fail in real call-center environments where speakers switch languages, mangle brand names, and misheard numbers are common. Prisma v3.0's Vistaar benchmark results specifically reflect its tuning for this messy, real-world multilingual speech rather than idealized dictation scenarios.
Further Reading
- Gnani AI launches Prisma v3.0, a speech-to-text model trained on Indian phone calls - The Hindu BusinessLine
- Gnani AI Prisma v3.0 Records 16.9% Word Error Rate - SMEStreet
- Gnani AI launches Prisma v3.0 speech-to-text for Indian calls - NewsBytes
- Prisma v3.0: ASR Model for Indian Languages - Gnani AI
