Introduction

Gnani AI launched Prisma v3.0 on October 28, a speech-to-text model built for the way India actually talks on the phone: Hindi, English and Kannada in one sentence, over a noisy street and a weak line. One model covers nine Indian languages plus English, figures out which one it's hearing, and lets you choose how every word gets written.

It's a closed model. You run it today through Gnani's developer API over REST, WebSocket streaming and batch, and Gnani also offers hybrid and fully on-premise deployments, including air-gapped ones. If you remember Gnani Timbre v3, the voice side of this family, Prisma is the ears.

What Shipped

From the launch post and the Prisma v3.0 docs:

  • Languages: Hindi, Bengali, Marathi, Gujarati, Punjabi, Kannada, Telugu, Tamil, Malayalam and English, all in one model.
  • Auto language detection: send indic-auto, or give one to four likely languages as a hint. Gnani reports 98% language-ID accuracy on clips longer than 3 seconds.
  • Audio: native 8 kHz telephony audio and 16 kHz app or VoIP audio, trained on six real telephony codecs.
  • Four output modes: Verbatim, Transcribe, Code-mix and Transliterate, picked per request.
  • Context hints: up to 200 keyterms per request, plus a topic note, up to four speaker names and the previous transcript chunk.
  • Find and replace: up to ten spelling rules per request, so a brand always comes back spelled your way.
  • Model ID: prisma-v3.0. It's opt-in, so existing integrations stay on Prisma v2.5 until you switch.

The Four Ways to Write the Same Call

This is the part I like most. The same audio can come back four different ways, depending on who reads it:

  • Verbatim keeps every filler word, repeat and false start, with numbers spelled out. It's built for compliance and audit records.
  • Transcribe gives clean native-script text with amounts, dates and percentages as digits. This is the one to feed a voice agent.
  • Code-mix writes each word in the script of the language it was spoken in, so English terms like EMI or refund stay in Latin letters next to Devanagari or Kannada.
  • Transliterate writes the whole call in Latin letters, Hinglish-style, without translating it.

Gnani's docs show one clip four ways. Verbatim returns मेरा अकाउंट नंबर एक दो तीन चार पाँच है, while Transcribe returns मेरा account नंबर 12345 है, and Transliterate gives you mera account number ek do teen char paanch hai.

Pro Tip

Send format every time. On Prisma v3.0, a request with a single hi-IN or en-IN language and no format now returns digits, while v2.5 defaulted to verbatim. If you're migrating, setting it explicitly keeps your old output stable.

The Numbers, With Caveats

These are Gnani's own scores on the public Vistaar benchmark, using its official scoring protocol:

  • Macro WER across nine Indian languages: 16.93 for Prisma v3.0, against 19.98 for ElevenLabs Scribe v2, 20.51 for Microsoft Azure STT and 26.57 for Deepgram Nova 3.
  • Hindi: 11.72 WER, against 14.47 for Scribe v2.
  • Malayalam, the hardest language in the set: 27.91 WER, against 29.81 for Scribe v2.
  • Keyterm recall: up from 70.9% to 90.3% when the right terms are supplied. Gnani tested this on public read speech, so treat it as a best case.
  • Hallucinated text on noise-only audio: down from 13.4% to 4.2% compared with the previous model.
  • Hindi-English calls with language hints: WER fell from 34.14 to 18.93 compared with the previous model.

It's a vendor benchmark, so test it on your own audio before you trust it. Azure still beat it on one of the seven Vistaar datasets, MUCS.

How to Try It Today

You'll need a Gnani API key. A short REST call with auto language detection looks like this:

curl --location "https://api.vachana.ai/stt/v3" \
  --header "X-API-Key-ID: <API_KEY>" \
  --form audio_file=@clip.wav \
  --form model=prisma-v3.0 \
  --form language_code=indic-auto \
  --form format=codemix \
  --form 'keyterms=["SmartSave+","NACH mandate"]'

The reply includes transcript and detected_language_code. For live audio, open the WebSocket at /stt/v3/stream with the header x-model: prisma-v3.0. For recordings longer than 30 seconds, use the Batch Jobs API, because REST caps each clip at 30 seconds.

Know the limits before you build. Prisma v3.0 has no word-level timestamps, doesn't translate, and only labels speakers in batch jobs. Assamese and other languages outside the ten aren't supported yet. Gnani's launch materials don't list a price, so check your account before you plan around cost.

Why It Matters for Audio Creators

Most transcription tools treat code-mixed speech as a mistake to clean up. Prisma v3.0 treats it as the normal case, and that's a big deal for Indian podcasters, YouTubers and voice-agent builders. A Hinglish interview can come back with Hindi in Devanagari and English in Latin, ready for subtitles. You can also get a full romanised transcript for viewers who don't read every script.

The keyterm feature helps creators too. Pass your show name, guest names and product terms, and they stop getting mangled across ten languages. Pair it with a TTS voice like Timbre v3 and you have both halves of an Indian-language voice agent from one vendor.

Conclusion

Gnani Prisma v3.0 is a closed, API-first speech-to-text model that treats India's code-mixed phone calls as the main job. It brings auto language detection across ten languages, four writing modes, 200 keyterms per request, and a vendor-reported 16.93 WER on Vistaar against ElevenLabs Scribe v2's 19.98. If your audio switches languages mid-sentence, this is worth a test run this week.

—Anabel ♡