Gnani.ai Launches Prisma v2.5, a Sovereign Voice AI Model for India
An Indian startup called Gnani.ai has made a new voice AI called Prisma v2.5. People call it a “sovereign” model. That means the AI is made, taught, and stored inside India, and Indians control it. Prisma v2.5 is a speech-to-text model. This means it listens to people talk and writes down what they say. Gnani.ai made it for Indian languages and accents. Many tools from other countries do this part badly.
The company’s founder and boss is Ganesh Gopalan. (A founder is the person who started the company.) Gnani.ai makes voice AI for big companies. It works most with banks, insurance, and healthcare. With Prisma v2.5, the company is trying hard to build India’s own voice tech. It does not want to depend only on tools from other countries.
What Prisma v2.5 can do
Prisma v2.5 can work in 12 languages. To teach it, the company used 14 million hours of speech. The data was “proprietary,” which means the company owns it. The speech was “Indic,” which means it came from languages of the Indian region.
Real Indian speech is messy, and Prisma v2.5 is built to handle it. It can deal with different ways of speaking, loud background noise, and code-switching. Code-switching is when a person mixes two languages in one sentence. For example, they may use English and Hindi together. It also understands short phrases, numbers, and alphanumeric strings (these are mixes of letters and numbers). And it gets named entities (these are names of people or places) and special work words too.
Benchmarks and specs
A benchmark is a standard test. People use it to see how well models do against each other. Gnani.ai tests Prisma with two scores: word error rate (WER) and character error rate (CER). Both scores count how many mistakes a model makes when it turns speech into text. A lower score is better. The company says Prisma does better than its rivals on these tests. The rivals are ElevenLabs, Sarvam AI, and Microsoft. The exact scores were not shared in the source.
| Spec | Prisma v2.5 (as reported) |
|---|---|
| Model type | Speech-to-text (voice AI) |
| Languages supported | 12 |
| Training data | 14 million hours of proprietary Indic speech |
| Accuracy metrics | Word error rate (WER), character error rate (CER) |
| Claimed to outperform | ElevenLabs, Sarvam AI, Microsoft |
| Hosting | Indian data centres via E2E Networks |
| Target industries | BFSI, insurance, healthcare |
Built and hosted in India
Prisma v2.5 is stored on Indian computers, or data centres, run by a company called E2E Networks. Keeping the model on local computers cuts the latency. Latency is the wait between when you ask the AI something and when it answers. Less latency means faster, smoother talking.
Storing it in India also keeps the data inside the country. This matters a lot to banks and hospitals. They must keep their customers’ private details safe. Gnani.ai also gets support from the Indian government’s IndiaAI Mission. This is a national plan to build India’s own AI.
Funding and earlier launches
In March 2026, Gnani.ai raised $10 million (about ₹94 crore). This came in a Series B round. A Series B is a later step of startup funding, used to help the business grow. Aavishkaar Capital led the round. An old backer, InfoEdge Ventures, also joined in.
The company will use the money to find new customers, enter new fields, and reach other countries. It will also spend on research and on hiring good people. Prisma is built on earlier work. Gnani.ai launched the Vachana speech-to-text model in December 2025 (taught with over 1 million hours of data). It also launched the Inya voice-first AI model in 2026.
| Funding detail | As reported |
|---|---|
| Round | Series B (March 2026) |
| Amount | $10 million (~₹94 crore) |
| Lead investor | Aavishkaar Capital |
| Existing backer | InfoEdge Ventures |
Why it matters (especially for India and founders)
India has hundreds of languages and accents. Most voice tools from other countries learn mainly from English and a few big languages. So they often fail with real Indian speech. A model made and stored in India can help the customers those tools miss.
For people who start companies, Prisma shows a clear path. Pick a problem that big global firms ignore. Then own the data for that problem. Sovereign AI also fits India’s bigger goal of controlling its own tech. More and more, Indians use voice to use apps and banks. So homegrown models like Prisma could win big deals with large companies.
FAQ
What is Prisma v2.5?
It is a speech-to-text voice AI from the Indian startup Gnani.ai. It turns spoken words into text, and it works in 12 languages.
What does “sovereign voice AI” mean?
It means the AI is made, taught, and stored inside India, and Indians control it. It does not depend on servers or models from other countries.
How much data was Prisma trained on?
Gnani.ai says it taught Prisma v2.5 with 14 million hours of its own Indic speech data.
Who funds Gnani.ai?
Gnani.ai raised $10 million (about ₹94 crore) in a Series B round in March 2026. Aavishkaar Capital led it, and InfoEdge Ventures, an old backer, joined too.
The takeaway
Prisma v2.5 is a bet that India needs voice AI made for its own languages, accents, and rules. Gnani.ai keeps the model on Indian computers and teaches it with local speech. This way, the company is claiming its spot in the sovereign AI space. Many Indian businesses talk to customers in many languages. For them, this could be a big, real win.
Source: Inc42
