Gnani AI Releases Prisma v3.0 Speech Recognition Model, Supporting 10 Languages
Gnani AI has released its automatic speech recognition model Prisma v3.0, supporting nine Indian languages including Hindi and Bengali as well as English, and capable of transcribing multilingual conversations with intra-sentence Hindi-English code-mixing. On the Vistaar benchmark, the model achieves a macro-average WER of 16.93, lower than ElevenLabs Scribe v2's 19.98, Microsoft Azure Speech-to-Text's 20.51, and Deepgram Nova 3's 26.57. It offers four transcription modes—Verbatim, Transcribe, Code-mix, and Transliterate—and supports up to 200 custom vocabulary terms per request, raising the recall rate for these terms from 70.9% to 90.3%.
Body
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.
The model supports Hindi, Bengali, Marathi, Gujarati, Punjabi, Kannada, Telugu, Tamil, Malayalam and English. According to the company, Prisma v3.0 recorded a macro Word Error Rate (WER) of 16.93 on the Vistaar benchmark, compared with 19.98 for ElevenLabs Scribe v2, 20.51 for Microsoft Azure Speech-to-Text and 26.57 for Deepgram Nova 3.
“We built it because a correct transcript still wasn’t a usable one,” Gnani AI said in its announcement, highlighting the challenges of making speech recognition useful for real-world conversations.
The model outperforms Prisma 2.5 and targets transcription errors involving mixed-language speech, spoken numbers, brand names and short responses that can affect the accuracy of enterprise voice applications.
Prisma v3.0 offers four transcription modes: Verbatim, Transcribe, Code-mix and Transliterate. These modes let users preserve fillers and repetitions, generate cleaner text, retain words in their original scripts, or write speech in English letters.
The model also supports custom vocabulary prompting, allowing developers to supply up to 200 terms per request, including product names and technical expressions. Gnani reported that recall of supplied terms increased from 70.9% to 90.3% without additional training or fine-tuning. The company described this result as an upper bound based on an evaluation using public read speech with the relevant terms supplied.
For language identification, Gnani reported 98% accuracy on audio clips longer than three seconds across the Vistaar and Svarah datasets. The company also said false text generated from noise-only audio fell from 13.4% to 4.2% compared with Prisma 2.5.
In its language-level comparison, Prisma v3.0 recorded lower WER than Microsoft Azure Speech-to-Text across all nine supported Indian languages. The aggregate Vistaar results, however, reflect a separate comparison across the benchmark's datasets.
The model accepts 8 kHz telephony audio and 16 kHz audio from applications and voice-over-internet-protocol (VoIP) systems. Gnani said it can be deployed through cloud, hybrid and on-premises infrastructure, including air-gapped environments.