MUMBAI: If you have ever tried getting a straight answer out of an Indian call centre, you know the linguistic obstacle course: a bit of Hinglish here, a dash of Tanglish there, all delivered over a scratchy 8 kHz phone line. For most speech-to-text bots, this verbal fruit salad is an absolute nightmare. Bengaluru-based frontier AI outfit Gnani AI thinks it has found the secret sauce with the launch of its latest speech model, Gnani Prisma v3.0.
Trained specifically on real-world Indian telephone conversations, Prisma v3.0 promises to clean up the babble. It supports nine Indian languages alongside Indian English from a single engine. More importantly, it gives teams five controls and four dedicated transcription modes to tweak their output, whether they want every single “um” and “ah” recorded for legal audits or a crisp summary that turns spoken numbers directly into hard figures.
According to Gnani AI, the model achieved a macro word error rate (WER) of 16.9 per cent across nine Indian languages under the official Vistaar scoring protocol. That beats ElevenLabs Scribe v2’s 19.98 per cent by roughly 12 per cent, with Prisma v3.0 scoring lower WER across all seven Vistaar datasets. On audio clips over three seconds, its automatic language identification nailed the spoken tongue more than 98 per cent of the time.
For Indian enterprises, the real kick lies in handling inverse text normalization. When a customer says “सव्वादोन कोटी” on a Marathi collection call or “ಎರಡೂವರೆ ಕೆಜಿ” on a Kannada helpline, standard speech tools usually spew out long-winded text. Prisma v3.0 automatically reformats them into clean business units like Rs 2.25 crore or 2.5 kg. Even colloquial time stamps like “साडेदहा” smoothly transform into a neat 10:30. Developers can also feed the model up to 200 custom keywords per request, ensuring niche product names and brand jargon stop getting lost in translation.
The mechanics are built around four primary modes:
- Verbatim: Holds onto every filler word and pause, giving compliance and legal crews a complete audit trail.
- Transcribe: Delivers clean, readable text in native scripts with dates, amounts, and figures automatically formatted into Indian digit grouping.
- Code-mix: Preserves the natural blend of languages, writing each word in its spoken script while keeping English terms intact.
- Transliterate: Converts the entire conversation into English script while keeping the original spoken words unchanged.
Gnani AI chief executive and co-founder Ganesh Gopalan pointed out that Indian business runs on phone calls that bounce between languages inside a single breath. Chief product and engineering officer Bharath Shankar added that code-mixed speech is the default setting across the country, making flexible formatting an absolute necessity rather than a fancy perk.
Building on technology that already powers 50 million daily voice interactions, Prisma v3.0 is live now. Developers can test-drive the model on the Gnani APIs playground or deploy it across production workflows using REST, WebSocket streaming, and batch processing.
In a market where one misheard figure can derail an entire loan reconciliation, Gnani AI is betting that making sense of India’s everyday phone chatter is the ultimate winning call.