Latest in AI
Gnani Prisma v3.0: Indian Speech-to-Text API, Modes and Pricing
Gnani's new transcription model lets teams choose how a multilingual call becomes text, which changes what happens after recognition.
Gnani AI introduced Prisma v3.0 on October 28, a speech-to-text model for nine Indian languages and English. It adds four output modes for code-mixed calls, where speakers switch languages within a sentence, alongside controls for vocabulary and language detection. The practical change is greater control over the transcript that downstream systems receive.
For a voice agent, recognizing the words is only part of the job. The result also needs to preserve the right names, format amounts correctly and avoid turning a mixed-language sentence into text that is difficult to search.
What Prisma v3.0 changes
Gnani's launch explanation positions v3.0 as the successor to Prisma v2.5. Its three main changes concern output formatting, custom vocabulary and language detection. It handles native 8 kHz telephone audio as well as 16 kHz app and VoIP audio.
The output mode is chosen per request. That matters because a record used to review a dispute and a transcript fed into a live assistant have different requirements.
| Mode | What the output preserves or changes |
|---|---|
| Verbatim | Keeps fillers and spoken numbers as words |
| Transcribe | Produces cleaner text, with numbers as digits and symbols |
| Code-mix | Keeps each language in its own script, including English terms |
| Transliterate | Writes the spoken words using English letters |
Transliteration does not translate the call into English. It changes the writing system while keeping the speaker's language. That distinction is important if another system assumes every Latin-script transcript contains English sentences.
Languages, API access and pricing
The Prisma v3.0 product page lists Hindi, Bengali, Marathi, Gujarati, Punjabi, Kannada, Telugu, Tamil, Malayalam and English. It offers REST, WebSocket streaming and asynchronous batch access, with all four output modes available across those paths.
Pricing is usage-based, charged per hour of transcribed audio, with the same pricing structure across REST, streaming and batch. The page does not publish a numeric hourly rate. Enterprise volume pricing requires a quote. It also lists on-premises deployment through the company.
Prisma v2.5 remains available for now. Existing users can test v3.0 by selecting the new model in their requests or the playground, rather than treating the announcement as an immediate forced migration.
What the benchmark does and does not show
Gnani reports a macro word error rate of 16.93 on Vistaar, compared with 19.98 for ElevenLabs Scribe v2. Lower is better. The difference is 3.05 percentage points, or approximately 15.3% lower relative word error rate.
That is a company-run evaluation, not an independent reproduction. It provides a reason to test the model, but does not establish which service will perform best on a particular support queue, accent mix or recording setup. An aggregate score can also hide the kinds of mistakes that matter most to a business.
A misspelled filler and a wrong repayment amount both contribute to transcription errors, but they have different consequences. Evaluation should therefore include the fields the application acts on, not only whether the transcript looks fluent.
The useful test is downstream
Consider a customer call containing an English product name inside a Hindi sentence. A readable transcript may still render the product phonetically, causing a search or reporting system to miss it. Preserving the intended script helps, but the pipeline still needs to verify that the correct product was recognized.
Choosing an output format is part of the application contract. A compliance archive may need the original hesitation and repetition. A CRM field may need a normalized amount. Sending one cleaned transcript to both consumers can lose information that one of them needed.
That is where Prisma v3.0 is worth evaluating. Teams can compare its modes against the same approved sample calls, check critical names and amounts, and then measure whether less cleanup is required. The meaningful improvement would be fewer fragile transformations between the conversation and the system that uses it.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help teams explain complex AI products through clear technical writing and launch stories that developers and founders can understand and share.