Gnani AI launches Prisma v3.0 ASR with 16.93 macro WER on Vistaar across 10 languages

Bengaluru-based Gnani AI has launched Prisma v3.0, an automatic speech recognition (ASR) model built for multilingual conversations, including speech that switches between Indian languages and English within a sentence. According to the company, it recorded a macro Word Error Rate (WER) of 16.93 on the Vistaar benchmark. Gnani's announcement was covered by Analytics India Magazine.

Languages and benchmark claims

Prisma v3.0 supports Hindi, Bengali, Marathi, Gujarati, Punjabi, Kannada, Telugu, Tamil, Malayalam and English.

Gnani's reported macro WER on Vistaar, with competitors' figures as given by the company:

Model	Macro WER (Vistaar)
Gnani Prisma v3.0	16.93
ElevenLabs Scribe v2	19.98
Microsoft Azure Speech-to-Text	20.51
Deepgram Nova 3	26.57

In its language-level comparison, Gnani says Prisma v3.0 recorded lower WER than Microsoft Azure Speech-to-Text across all nine supported Indian languages. The aggregate Vistaar results reflect a separate comparison across the benchmark's datasets. All figures are company-reported.

Gnani also says the model outperforms its predecessor, Prisma 2.5. It targets errors involving mixed-language speech, spoken numbers, brand names and short responses, which can affect enterprise voice applications. "We built it because a correct transcript still wasn't a usable one," the company said in its announcement.

Transcription modes and features

Prisma v3.0 offers four transcription modes:

• Verbatim: preserves fillers and repetitions.
• Transcribe: generates cleaner text.
• Code-mix: retains words in their original scripts.
• Transliterate: writes speech in English letters.

The model supports custom vocabulary prompting, with up to 200 terms per request, such as product names and technical expressions. Gnani reports that recall of supplied terms rose from 70.9% to 90.3% without additional training or fine-tuning. The company describes this as an upper bound, based on an evaluation using public read speech with the relevant terms supplied.

For language identification, Gnani reports 98% accuracy on audio clips longer than three seconds across the Vistaar and Svarah datasets. It also says false text generated from noise-only audio fell from 13.4% to 4.2% compared with Prisma 2.5.

Audio input and deployment

The model accepts 8 kHz telephony audio and 16 kHz audio from applications and voice-over-internet-protocol (VoIP) systems. Gnani says it can be deployed through cloud, hybrid and on-premises infrastructure, including air-gapped environments. Pricing is not yet disclosed. Parameter count and training data details are also not yet disclosed.

What this means

The headline WER numbers are Gnani's own, so they describe the company's evaluation rather than an independent test. The 90.3% term-recall figure carries the company's own caveat: it is an upper bound from read speech with the terms supplied, not a measure of live conversations.

The feature set points at practical transcription problems in Indian-language voice systems: code-mixed speech, spoken numbers, brand names and noise-only audio. The 8 kHz telephony support and air-gapped deployment options matter for call-based and restricted-network use. The four output modes let developers choose between faithful and cleaned-up transcripts, and between native scripts and English letters, without switching models.