Gnani AI Targets Mixed-Language Speech With Prisma v3.0

Gnani AI has launched Prisma v3.0, an automatic speech recognition model built for multilingual conversations across ten languages. The model supports Hindi, Bengali, Marathi, Gujarati, Punjabi, Kannada, Telugu, Tamil, Malayalam and English, including conversations that switch between languages within a single sentence.

That focus on mixed-language speech sits at the center of the new model. Gnani AI said it created Prisma v3.0 because “a correct transcript still wasn’t a usable one.” The model targets errors involving code-switching, spoken numbers, brand names and short responses, all of which can affect enterprise voice applications.

On the Vistaar benchmark, Prisma v3.0 recorded a macro Word Error Rate, or WER, of 16.93. The comparison figures were 19.98 for ElevenLabs Scribe v2, 20.51 for Microsoft Azure Speech-to-Text and 26.57 for Deepgram Nova 3.

Four ways to handle a conversation

Prisma v3.0 offers four transcription modes: Verbatim, Transcribe, Code-mix and Transliterate. Each mode is designed for a different way of working with spoken language, giving users control over how the final transcript represents a conversation.

• Verbatim: Preserves fillers and repetitions in speech.
• Transcribe: Produces cleaner text.
• Code-mix: Retains words in their original scripts.
• Transliterate: Writes speech in English letters.

This setup separates a word’s spoken content from the way a user wants that content displayed. A transcript can preserve the details of a conversation, create cleaner text, retain original scripts or represent speech through English letters.

The model also supports custom vocabulary prompting. Developers can supply up to 200 terms per request, including product names and technical expressions. Gnani AI reported that recall of supplied terms rose from 70.9% to 90.3% without additional training or fine-tuning.

Gnani AI described that result as an upper bound based on an evaluation using public read speech with the relevant terms supplied. The figure focuses on how often the model recognized the provided vocabulary, which can matter when a transcript includes specialized terms.

Performance across languages and audio types

Prisma v3.0 recorded lower WER than Microsoft Azure Speech-to-Text across all nine supported Indian languages in Gnani AI’s language-level comparison. The aggregate Vistaar results represent a separate comparison across the benchmark’s datasets, so the two comparisons measure different views of the model’s performance.

For language identification, Gnani AI reported 98% accuracy on audio clips longer than three seconds across the Vistaar and Svarah datasets. The model also reduced false text generated from noise-only audio from 13.4% with Prisma 2.5 to 4.2% with Prisma v3.0.

These results connect to the model’s focus on real speech conditions, including audio that contains more than one language and recordings with no spoken words. Prisma v3.0 outperforms Prisma 2.5 by targeting the transcription errors that Gnani AI identified in mixed-language speech and enterprise applications.

The model accepts 8 kHz telephony audio as well as 16 kHz audio from applications and voice-over-internet-protocol, or VoIP, systems. That gives it support for the two audio formats listed by Gnani AI for phone-based and application-based speech.

Deployment options for organizations

Gnani AI said Prisma v3.0 can be deployed through cloud, hybrid and on-premises infrastructure, including air-gapped environments. The deployment choices cover organizations using cloud systems, a mix of infrastructure or local environments separated from outside networks.

That range matches the model’s enterprise focus. Prisma v3.0 combines multilingual transcription, code-switching support, custom vocabulary prompting and four output modes in one speech recognition model, while its reported results compare it with Prisma 2.5 and three competing systems.

For developers, the model’s vocabulary controls offer a way to supply product names and technical expressions at request time rather than using additional training or fine-tuning. For users handling multilingual speech, the four modes provide different ways to preserve, clean, script or represent the words in a transcript.

Stay connected via Google News

Follow us for the latest AI News