Bengaluru-based artificial intelligence startup Gnani AI has launched Prisma v3.0, a new automatic speech recognition (ASR) model designed to transcribe conversations across 10 languages, including nine Indian languages and English. The model focuses on a major challenge in India’s voice technology market: accurately understanding conversations in which people frequently switch between languages within the same sentence.
According to Gnani AI, Prisma v3.0 achieved a macro Word Error Rate (WER) of 16.93 on the Vistaar benchmark, outperforming several competing speech recognition systems in the company’s evaluation. The model introduces four transcription modes, custom vocabulary support and improved language identification, targeting enterprise applications such as customer service, banking, sales and automated voice agents.
Key Takeaways
- 10-language support: Prisma v3.0 supports Hindi, Bengali, Marathi, Gujarati, Punjabi, Kannada, Telugu, Tamil, Malayalam and English.
- 16.93 Word Error Rate: The model achieved a macro WER of 16.93 on the Vistaar benchmark, according to Gnani AI.
- Four transcription modes: Users can choose Verbatim, Transcribe, Code-mix and Transliterate.
- 98% language identification: Gnani reports 98% accuracy for identifying languages in audio clips longer than three seconds.
- Custom vocabulary: Developers can provide up to 200 specialised terms per request to improve recognition of brand names and technical expressions.
- Reduced transcription errors: False text generated from noise-only audio declined from 13.4% to 4.2% compared with Prisma 2.5.
- Enterprise deployment: The model supports cloud, hybrid and on-premises deployment, including air-gapped environments.
What Is Gnani AI’s Prisma v3.0?
Prisma v3.0 is an automatic speech recognition model developed by Gnani AI to convert spoken conversations into written text.
Automatic speech recognition, commonly called ASR, is the technology that allows computers to understand spoken language and generate transcripts.
It is widely used in customer support systems, call centres, voice assistants, meeting transcription tools and automated telephone services.
However, recognising Indian conversations can be particularly challenging because people frequently combine English with regional languages.
For example, someone might say, “Sir, mera payment update nahi hua” while switching between English and Hindi.
A conventional transcription system may recognise the sentence incorrectly, write English words in the wrong script or misunderstand important details such as amounts, product names and customer instructions.
Prisma v3.0 is designed to address these challenges by improving multilingual recognition and providing greater control over how spoken conversations are converted into text.
The model supports nine Indian languages alongside English, making it relevant for businesses handling multilingual customer interactions.
Prisma v3.0 Supports 10 Languages
One of the main features of Prisma v3.0 is its multilingual speech recognition capability.
The model supports the following languages:
| No. | Supported language |
|---|---|
| 1 | Hindi |
| 2 | Bengali |
| 3 | Marathi |
| 4 | Gujarati |
| 5 | Punjabi |
| 6 | Kannada |
| 7 | Telugu |
| 8 | Tamil |
| 9 | Malayalam |
| 10 | English |
The model is designed to recognise conversations involving multiple languages rather than requiring speakers to communicate in only one language.
This is particularly important for Indian businesses operating across different states, where customers may switch between regional languages and English during telephone conversations.
For example, a customer speaking in Hindi might use English terms such as EMI, refund, account balance or transaction ID.
Prisma v3.0 aims to preserve these expressions accurately so that downstream software systems can understand the conversation.
Prisma v3.0 Records 16.93 Word Error Rate
Another important part of the launch is the model’s reported performance on the Vistaar benchmark.
Word Error Rate (WER) is a standard metric used to measure the accuracy of speech recognition systems.
It calculates errors involving:
- Substitutions: When the model replaces a spoken word with an incorrect word.
- Deletions: When the model misses a spoken word.
- Insertions: When the model generates words that were never spoken.
A lower WER indicates better transcription accuracy.
According to Gnani AI, Prisma v3.0 achieved a macro WER of 16.93 across nine Indian languages and seven Vistaar datasets.
The company compared its performance with several established speech recognition systems.
| Speech recognition model | Macro WER |
|---|---|
| Gnani Prisma v3.0 | 16.93 |
| ElevenLabs Scribe v2 | 19.98 |
| Microsoft Azure Speech-to-Text | 20.51 |
| Deepgram Nova 3 | 26.57 |
<text color=”secondary” size=”xs”>Source: Gnani AI’s published Vistaar benchmark evaluation. Lower WER is better. These are company-reported comparisons, not an independent verification of overall product performance.</text>
The results suggest that Prisma v3.0 can provide more accurate transcription than the compared systems under the evaluated conditions.
However, performance can vary depending on recording quality, accents, background noise and the language being spoken.
Four Transcription Modes Give Developers More Control
Prisma v3.0 introduces four different transcription modes designed for different business requirements.
Instead of generating the same type of transcript for every application, developers can select how the model should format the output.
1. Verbatim Mode
Verbatim mode preserves conversations as closely as possible to the original speech.
It retains filler words, repetitions and incomplete sentences.
This is useful for:
- Legal and compliance records.
- Customer complaint investigations.
- Call quality monitoring.
- Audit trails.
- Employee performance evaluation.
2. Transcribe Mode
Transcribe mode produces cleaner and more readable text.
It removes unnecessary repetitions and converts spoken numbers, dates and percentages into written numerical formats.
For example, spoken expressions can be converted into figures rather than remaining as words.
This is useful for customer relationship management systems, call summaries and automated customer support.
3. Code-mix Mode
Code-mix mode is particularly relevant for Indian conversations.
It preserves each word in the script of the language in which it was spoken.
For example, a sentence combining Hindi and English can retain Hindi words in Devanagari while keeping English terms in Latin characters.
This can improve the usefulness of transcripts for multilingual voice agents and search systems.
4. Transliterate Mode
Transliterate mode converts spoken regional-language words into English letters without translating their meaning.
For example, Hindi speech can be written using Roman characters instead of Devanagari.
This is useful for businesses whose internal systems primarily support English-language interfaces.
Custom Vocabulary Helps Recognise Brand Names and Technical Terms
Another important feature is custom vocabulary prompting.
Speech recognition systems sometimes struggle with specialised terminology, company names, product names and industry-specific expressions.
For example, a customer mentioning a particular banking product or technical service may have their words incorrectly transcribed.
Prisma v3.0 allows developers to provide up to 200 custom terms per request.
These can include:
- Product and brand names.
- Banking terminology.
- Technical expressions.
- Company names.
- Specialised industry vocabulary.
Gnani AI reported that recall of supplied terms increased from 70.9% to 90.3% when custom vocabulary prompting was used.
The company described this result as an upper bound based on an evaluation using public read speech with the relevant terms supplied.
The improvement was achieved without additional model training or fine-tuning.
This feature could be particularly useful for banks, insurance companies, ecommerce businesses and customer service providers that regularly encounter specialised vocabulary.
Language Identification Accuracy Reaches 98%
Prisma v3.0 also includes automatic language identification.
This feature allows the model to identify the language being spoken without requiring users to manually select it before every conversation.
According to Gnani AI, the system achieved 98% language identification accuracy on audio clips longer than three seconds across the Vistaar and Svarah datasets.
This capability could help businesses operate multilingual voice agents without creating separate systems for every supported language.
The model also supports language hints and candidate-language lists, allowing developers to guide language recognition when they know which languages are likely to appear in a conversation.
Designed for Noisy Telephone Conversations
Many enterprise voice applications operate through traditional telephone infrastructure.
These systems often use lower-quality audio than modern smartphone recordings.
Prisma v3.0 supports native 8 kHz telephony audio and 16 kHz audio used by applications and Voice over Internet Protocol (VoIP) systems.
Gnani AI also reported that false text generated from noise-only audio declined from 13.4% with Prisma 2.5 to 4.2% with Prisma v3.0.
This improvement matters because background noise and poor-quality recordings can cause speech recognition systems to generate words that were never actually spoken.
The model is designed to support real-world business conversations rather than only clean recordings made in controlled environments.
Where Can Businesses Use Prisma v3.0?
Gnani AI is targeting enterprise applications where accurate multilingual transcription can improve business operations.
Potential applications include:
| Industry | Potential application |
|---|---|
| Banking | Transcribing customer calls and financial instructions |
| Insurance | Recording customer conversations and claims |
| Ecommerce | Understanding customer queries and complaints |
| Healthcare | Transcribing patient conversations |
| Telecommunications | Analysing customer support calls |
| Retail | Understanding customer feedback |
| Education | Transcribing multilingual lessons |
| Customer support | Improving automated voice agents |
The ability to recognise regional languages and mixed-language conversations could help businesses serve customers who are more comfortable communicating in their native languages.
The model also supports cloud, hybrid and on-premises deployment, including air-gapped environments.
This gives organisations additional flexibility when handling sensitive conversations and internal business information.
How Does Prisma v3.0 Compare With Gnani’s Previous Model?
Prisma v3.0 builds on Gnani AI’s earlier Prisma 2.5 speech recognition system.
The latest version focuses on improving several practical transcription problems, including code-mixed speech, language identification, custom terminology and noise handling.
| Feature | Prisma 2.5 | Prisma v3.0 |
|---|---|---|
| Language support | Earlier multilingual capabilities | 10 languages |
| Transcription modes | Earlier output capabilities | Four modes |
| Custom vocabulary | Earlier capabilities | Up to 200 terms per request |
| Language identification | Previous performance | 98% reported accuracy on clips over 3 seconds |
| False text on noise-only audio | 13.4% | 4.2% |
| Deployment | Existing enterprise options | Cloud, hybrid and on-premises, including air-gapped environments |
Gnani AI’s latest release indicates that the company is focusing on making speech recognition more useful for real business workflows rather than concentrating only on headline accuracy scores.
The Bigger Picture
Gnani AI’s Prisma v3.0 launch highlights the growing importance of Indian-language artificial intelligence.
India has a large multilingual population, and businesses increasingly need technology that can understand customers communicating in different regional languages.
Speech recognition is an important foundation for this transition because voice agents, automated customer support and conversational AI systems depend on accurate transcripts.
The competition is also expanding. Global companies such as Microsoft, ElevenLabs and Deepgram offer speech recognition technologies, while Indian AI companies are developing systems designed around local languages and communication patterns.
Prisma v3.0’s focus on code-mixed speech, telephone audio and regional-language accuracy reflects an effort to address requirements that are particularly relevant to Indian enterprises.
Looking Ahead
Gnani AI’s next challenge will be translating its benchmark results into reliable performance across real customer conversations. The model’s support for multiple Indian languages, custom vocabulary and flexible deployment could make it relevant to businesses building multilingual voice agents and automated customer service systems.
As companies increasingly adopt conversational AI, speech recognition models designed for regional languages could become an important part of India’s AI infrastructure. Prisma v3.0 adds another competitor to this growing market, although broader adoption will depend on accuracy, pricing, integration and performance at enterprise scale.
Frequently Asked Questions
1. What is Gnani AI Prisma v3.0?
Prisma v3.0 is an automatic speech recognition model developed by Gnani AI. It converts spoken conversations into text and supports nine Indian languages and English.
2. Which languages does Prisma v3.0 support?
It supports Hindi, Bengali, Marathi, Gujarati, Punjabi, Kannada, Telugu, Tamil, Malayalam and English.
3. What is the Word Error Rate of Prisma v3.0?
Gnani AI reported a macro Word Error Rate of 16.93 on the Vistaar benchmark. Lower WER indicates better transcription accuracy.
4. What are the four transcription modes?
The four modes are Verbatim, Transcribe, Code-mix and Transliterate. They provide different ways to format spoken conversations.
5. Can Prisma v3.0 understand Hinglish?
Yes. The model is designed to recognise conversations that combine Hindi and English within the same sentence. Its Code-mix mode preserves words in their respective scripts.
6. Can businesses use Prisma v3.0 for customer service?
Yes. The model is designed for enterprise applications, including customer support, banking, insurance and automated voice agents.
7. Does Prisma v3.0 support on-premises deployment?
Yes. Gnani AI says the model supports cloud, hybrid and on-premises deployment, including air-gapped environments.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
