Gnani.ai, a Bengaluru-based enterprise voice AI startup, released Prisma v2.5, a speech-to-text model supporting 12 languages, on June 17, according to an Inc42 report. The model was trained on 14 million hours of proprietary Indic speech data and is now available to enterprise customers via APIs.

What the model does

Prisma v2.5 is designed to transcribe Indian-language speech, accounting for dialect variation, background noise, and mid-sentence code-switching. According to Gnani.ai, these factors are integrated into the training data rather than treated as exceptions.

The company states that the model improves accuracy for short utterances, numbers, alphanumeric strings, and named entities. Errors in these areas often lead to compliance, CRM, and customer service issues across sectors such as banking, insurance, and healthcare.

Gnani.ai co-founder and CEO Ganesh Gopalan explained that most automatic speech recognition models are designed for studio-quality audio. In contrast, Indian phone calls often feature compressed network audio, multiple languages in a single sentence, and diverse accents not covered by standard datasets.

Performance claims

Gnani.ai reports that it benchmarked Prisma on word and character error rates, with both internal and third-party evaluations indicating it outperforms ElevenLabs, Sarvam AI, and Microsoft on these metrics. However, the company did not disclose its benchmark methodology or specific scores in the announcement.

Gopalan stated that an early retail client switched from a global STT provider to Prisma, though the client was not named.

The company attributes the lower latency to hosting the model in Indian data centres, including E2E Networks’ infrastructure. It says this approach makes the model better suited to real-time applications such as telephony than models hosted overseas.

Rollout plans beyond India

The current release supports only Indian languages. Gopalan stated that Gnani.ai plans to expand the model to Japan, the Philippines, and the Middle East, aiming to establish Prisma as a global product.

Prisma v2.5 builds on two previous releases: Vachana STT, an enterprise Indic speech-to-text model trained on over 1 million hours of voice data and launched in December 2025, and Inya, a voice-first AI model introduced at the India Impact AI Summit 2026.

Gnani.ai raised $10 million (approximately Rs 94 crore) in a Series B round in March, led by Aavishkaar Capital with participation from existing investor InfoEdge Ventures. The company stated that the funds will support expansion into new verticals and geographies, as well as research, development, and hiring.

Competitive backdrop

The launch comes as competition intensifies in India’s voice AI sector between domestic “sovereign AI” companies and global firms such as ElevenLabs and Wispr Flow. Gnani.ai, along with Sarvam AI, Fractal Analytics, and BharatGen, is developing models in this space. Gnani.ai is also among the companies supported by the government’s IndiaAI Mission, together with Sarvam and Soket AI.

Gopalan stated that Gnani.ai’s models are more localised and tested under real-world Indian conditions, which he believes provides a competitive advantage in accuracy.

The sovereign AI discussion has intensified after the United States last week imposed export-control restrictions on Anthropic’s Fable 5 and Mythos 5 models, citing national security concerns. As a result, Anthropic disabled the models globally instead of implementing selective access controls. Gopalan noted that this incident highlights the importance of countries developing their own AI models with independent security safeguards, a point he has previously emphasised, but one that is now more urgent.

Also read:

• What Are the Risks of Sarvam AI’s New Text-to-Speech Model for Indian Users?
• Explained: OpenAI’s Efforts For Audio-First Personal Device And What It Tells About Voice-First AI
• Microsoft Launches New Speech Generation AI Model MAI-Voice-1 Despite AI Capacity Constraints