Gnani AI has launched Prisma v3.0, a speech-to-text model designed to transcribe Indian phone calls with greater accuracy across regional languages, accents and code-switching. The Bengaluru-based AI company says the model supports nine Indian languages and Indian English, with four transcription modes that let businesses choose how spoken conversations are converted into text.
Announced on October 28, 2026, Prisma v3.0 targets a common problem for Indian enterprises: conversations often move between languages within the same sentence, while telephone audio can be noisy or compressed. The model is designed to turn these conversations into usable records for customer support, banking, insurance, sales and compliance workflows. Gnani AI also reports improved results against competing speech recognition systems in its Vistaar benchmark evaluation.
Key Takeaways
- New AI model: Gnani AI launched Prisma v3.0 on October 28, 2026.
- Language support: The model supports nine Indian languages plus Indian English.
- Four transcription modes: Verbatim, Transcribe, Code-mix and Transliterate serve different business requirements.
- Benchmark results: Gnani reports a 16.9% macro word error rate across nine Indian languages, compared with 19.98% for ElevenLabs Scribe v2.
- Language identification: Automatic language identification achieved more than 98% accuracy on Vistaar audio clips longer than three seconds, according to the company.
- Developer access: Prisma v3.0 is available through REST APIs, WebSocket streaming and batch processing.
What Is Gnani AI’s Prisma v3.0?
Prisma v3.0 is a speech-to-text model built to convert spoken conversations into text that businesses can use in their software systems.
Traditional speech recognition can struggle when speakers switch languages, pronounce brand names differently or communicate over low-quality telephone connections. A conversation might begin in Hindi, include English business terms and finish in Kannada or another regional language. A transcription system that assumes only one language may misinterpret words or produce text that requires substantial manual correction.
Gnani AI says Prisma v3.0 is trained on Indian telephone audio and heavily code-mixed speech. This focus is intended to improve transcription in real-world business environments rather than only on clean recordings or formal speech.
The company says its speech models already power approximately 50 million voice interactions daily. Prisma v3.0 extends that technology with greater control over how the same audio is represented in the final transcript.
Four Transcription Modes for Different Business Needs
One of Prisma v3.0’s main features is the ability to produce different transcript styles from the same recording. Businesses can select a mode for each request depending on whether they need an exact record, readable text or compatibility with existing software.
| Mode | How it works | Typical use |
|---|---|---|
| Verbatim | Preserves fillers, repetitions and spoken wording | Compliance, audits and dispute investigations |
| Transcribe | Produces cleaner text with numbers and amounts converted into digits | Customer-service records and operational workflows |
| Code-mix | Preserves each language in its own script while retaining English terms in English | Multilingual CRM records and call analytics |
| Transliterate | Writes the spoken words using English letters | Systems that work better with Roman-script text |
The four options address a practical limitation in speech recognition: there is no single transcript format that works equally well for every department.
A compliance team may need to preserve exactly what a customer said, including repetitions and hesitation. A customer-service team, by contrast, may prefer a clean transcript that can be added directly to a customer relationship management system.
Code-mix is particularly relevant in India, where speakers frequently combine languages in a single conversation. Transliteration provides another option for teams that need the spoken language represented in English letters rather than its native script.
Prisma v3.0 Benchmark Results Compared With Competitors
Gnani AI evaluated Prisma v3.0 using the Vistaar scoring protocol, reporting a macro word error rate of 16.9% across nine Indian languages. The company compared that result with 19.98% for ElevenLabs Scribe v2.
| Evaluation metric | Prisma v3.0 | ElevenLabs Scribe v2 |
|---|---|---|
| Macro word error rate across nine Indian languages | 16.9% | 19.98% |
| Relative difference in reported WER | Lower | Higher |
| Automatic language identification on Vistaar clips longer than 3 seconds | More than 98% accuracy reported | No directly comparable figure provided in the cited launch report |
Word error rate, or WER, measures transcription errors involving substitutions, deletions and insertions. A lower score generally indicates better transcription performance.
Using the figures reported by Gnani AI, Prisma v3.0’s WER is approximately 15.4% lower relative to the 19.98% baseline—not a 15.4-percentage-point reduction. The company also says Prisma v3.0 achieved a lower WER on each of the seven Vistaar datasets used in the comparison.
These are company-reported benchmark results, not an independent guarantee of superior performance for every language or recording. Real-world accuracy can vary with background noise, accents, speaker overlap, audio quality and the vocabulary used during a call. Enterprises should test the model on representative recordings before choosing it for critical workflows.
Automatic Language Identification and Number Formatting
Prisma v3.0 includes automatic language identification, designed to recognize which language is being spoken without requiring the developer to specify every language switch in advance.
Gnani AI says the feature correctly identified the spoken language more than 98% of the time on Vistaar audio clips longer than three seconds. The company says the capability is intended to support conversations in which speakers switch languages mid-sentence.
The model also offers inverse text normalization, which converts spoken numbers into standard numeric representations. For example, a spoken amount such as “three lakh fifty thousand” can be represented as “3,50,000” in supported transcription modes.
This can reduce manual work for organizations processing loan discussions, payment queries, insurance claims or sales calls. Correctly formatted numbers can be easier to search, compare and transfer into business systems.
Developers can also supply up to 200 keywords per request, helping the model recognize specific product names, company terminology and other important vocabulary.
How Businesses Can Use Prisma v3.0
The model is designed for workflows in which spoken conversations need to become searchable, structured business information.
Banking and financial services: Lenders can transcribe customer discussions about repayments, loan terms and payment disputes. Verbatim records may help with audits, while cleaner transcripts can support operational processes.
Customer support and contact centers: Support teams can generate transcripts for call summaries, quality monitoring and customer relationship management records. Streaming transcription can also support applications that respond while a conversation is taking place.
Insurance and healthcare administration: Organizations can use transcription to organize customer conversations, subject to applicable privacy and regulatory requirements. Accuracy remains especially important when records contain policy details, names or financial information.
Sales and retail: Multilingual transcripts can help teams search conversations for product names, customer requests and frequently discussed issues.
The model does not automatically guarantee accurate business decisions or regulatory compliance. Organizations must still review sensitive records, protect customer information and establish quality controls appropriate to the intended use.
API Availability and Integration Options
Gnani AI says Prisma v3.0 is available through its developer API, with three integration methods:
- REST API: Suitable for sending recordings or short audio clips and receiving transcripts in response.
- WebSocket streaming: Designed for live audio and applications that need transcription as the conversation unfolds.
- Batch processing: Intended for large volumes of existing recordings that can be processed asynchronously.
These options let developers integrate the same speech recognition technology into real-time voice agents, recorded-call analysis and large-scale quality assurance systems.
The company also provides a developer platform and playground for testing the model. Pricing is usage-based, according to its product documentation, with enterprise arrangements available through its sales team.
The Bigger Picture
Prisma v3.0 reflects a broader trend in enterprise AI: models are increasingly being designed for the languages, audio conditions and workflows of specific markets. For India, that means handling not only regional languages but also the way people combine those languages with English in everyday business conversations.
The commercial value of speech recognition depends on more than converting sound into words. Businesses need transcripts that preserve evidence when necessary, format numbers consistently and fit into existing software. By offering multiple transcription modes, language identification and keyword prompting, Gnani AI is targeting these operational requirements.
The competitive test will be whether the model delivers reliable results across different industries, accents and recording conditions. If its reported benchmark advantages translate into customer deployments, Prisma v3.0 could help reduce the manual effort required to turn India’s multilingual phone conversations into usable business data.
FAQs
1. What is Gnani AI Prisma v3.0?
Prisma v3.0 is a speech-to-text model launched by Gnani AI on October 28, 2026. It is designed for Indian telephone audio, regional languages and code-mixed conversations.
2. Which languages does Prisma v3.0 support?
The model supports nine Indian languages—Hindi, Bengali, Marathi, Gujarati, Punjabi, Kannada, Telugu, Tamil and Malayalam—along with Indian English.
3. What are the four transcription modes?
Verbatim preserves spoken wording and repetitions. Transcribe produces cleaner text with normalized numbers. Code-mix preserves the scripts of different languages in the same conversation. Transliterate represents spoken words using English letters.
4. How accurate is Prisma v3.0?
Gnani AI reports a 16.9% macro word error rate across nine Indian languages on the Vistaar evaluation, compared with 19.98% for ElevenLabs Scribe v2. It also reports more than 98% language-identification accuracy on Vistaar clips longer than three seconds. These results are specific to the company’s stated evaluation conditions.
5. Can developers integrate Prisma v3.0 into applications?
Yes. Gnani AI offers REST, WebSocket streaming and batch API access. These options support recorded audio, real-time voice applications and large-scale call-processing workflows.
Looking Ahead
Prisma v3.0 will be worth watching as Indian businesses expand their use of voice agents, automated call summaries and AI-powered contact-center tools. The model’s combination of code-mixed transcription and configurable output formats addresses a real challenge for organizations serving multilingual customers.
The next measure of success will be performance outside benchmark datasets. Enterprises will need to assess transcription quality, language identification, processing speed, pricing and data-handling safeguards on their own recordings. Those results will determine how effectively Prisma v3.0 can turn India’s everyday business calls into reliable, actionable information.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
