Indian artificial intelligence company Sarvam AI has unveiled Saaras V4, its latest automatic speech recognition (ASR) model designed to handle multilingual speech, noisy real-world audio, code-mixing and multiple speakers.
The new model is aimed at making speech-to-text systems more useful across India’s diverse language landscape while also improving performance with global English accents and difficult audio conditions. Sarvam says Saaras V4 can process speech in 22 Indian languages plus English and automatically identify the spoken language before producing a transcript in the appropriate script.
One of the model’s more notable capabilities is its support for five different output formats from the same audio input. Developers can generate standard transcription, verbatim text, code-mixed text, transliteration or translation without building separate processing pipelines for each format.
What Is Sarvam AI Saaras V4?
Saaras V4 is an automatic speech recognition model developed by Sarvam AI for converting spoken audio into text.
Unlike speech recognition systems designed primarily around clean recordings, Sarvam says Saaras V4 has been built for real-world Indian speech, where conversations can contain background noise, different accents, regional dialects and a mixture of Indian languages and English.
The model combines an audio encoder with a 3-billion-parameter hybrid state-space language model that Sarvam trained in-house from scratch.
This architecture is intended to help the system understand both the acoustic characteristics of speech and the linguistic context needed to produce more useful transcripts.
According to Sarvam, Saaras V4 also extends English recognition beyond Indian English to include global English accents.
Saaras V4 Supports 22 Indian Languages
Language diversity is one of the main areas where Sarvam is positioning Saaras V4.
The model supports 22 Indian languages along with English, covering languages such as Hindi, Bengali, Marathi, Tamil, Telugu, Kannada, Malayalam, Gujarati, Punjabi, Assamese, Odia and Urdu.
It can also automatically detect the language being spoken.
This means developers do not necessarily need to specify the language before processing an audio recording. Sarvam says its language identification error rate is 5.22% across all 22 Indian languages, falling to 2.9% across the 10 most widely spoken Indian languages on its reported evaluation data.
Once the language has been identified, Saaras V4 can produce text in the corresponding native script.
For multilingual applications, this can simplify the development process because a single speech recognition model can handle several languages instead of requiring separate models or manually selected language configurations.
Five Output Modes From the Same Audio
Another major feature of Saaras V4 is its ability to produce different forms of text from the same spoken input.
Sarvam provides five output modes:
- Transcribe: Standard transcription with formatting and number normalization
- Verbatim: Word-for-word transcription that preserves spoken fillers and original forms
- Codemix: Keeps English words in English while retaining Indic-language text in its native script
- Translit: Converts speech into Roman or Latin script
- Translate: Converts supported Indic-language speech into English
These modes are built into the model rather than requiring separate post-processing systems, according to Sarvam.
This could be useful for different types of applications.
For example, a customer-support platform could use normalized transcription for records, while a consumer application might use transliteration to display Indian-language speech in the Latin alphabet. Developers could also use translation when they need English output from an Indic-language recording.
Saaras V4 Is Designed for Noisy Audio
Real-world audio is rarely as clean as recordings used in controlled speech-recognition tests.
Background conversations, traffic, wind, distortion, poor microphones and compressed telephone audio can all reduce transcription accuracy.
Sarvam says Saaras V4 has specifically been designed to handle these conditions. The company reports that the model is robust to noisy audio, dialectal variation and code-mixed speech.
Sarvam’s examples include Hindi recordings with wind noise and distortion, Bengali audio recorded in heavy traffic and Kannada speech containing static noise.
This focus could make the model relevant for customer-service calls, field recordings, meetings, voice assistants and other applications where audio quality cannot always be controlled.
Multi-Source Audio Processing
Saaras V4 is also designed to handle situations involving multiple audio sources.
According to NDTV Profit’s report, the model can listen to multiple audio tracks simultaneously and track different sources in a single pass.
That capability could be useful for applications where multiple speakers or audio channels need to be processed together.
For example, businesses could use speech recognition for meetings, interviews or customer-service recordings where several voices are present.
Sarvam also supports speaker diarization through its batch API, allowing multi-speaker conversations to be processed with speaker-specific timestamps.
Saaras V4 Performance Across English Benchmarks
Sarvam says it evaluated Saaras V4 across seven English speech-recognition benchmarks covering different conditions.
The evaluation included Indian English, international accents, meetings, media, financial conversations and other speech environments. Sarvam reports that Saaras V4 achieved the lowest average Word Error Rate across these datasets.
The benchmarks included AMI for meeting-room speech, GigaSpeech for podcasts and web video, LibriSpeech, SPGISpeech for financial vocabulary, VoxPopuli and Svarah for Indian-accented English.
These results are company-reported benchmark results, so they should be interpreted in the context of Sarvam’s stated evaluation methodology rather than treated as an independent universal ranking of speech-recognition models.
Performance on Indian Language Benchmarks
For Indian languages, Saaras V4 was tested on Sarvam’s Vistaar evaluation across 10 languages.
The evaluation used both conventional Word Error Rate (WER) and LLM-WER. The latter attempts to account for semantic differences and distinguish meaningful transcription mistakes from differences involving formatting or orthography.
Sarvam also reports state-of-the-art performance across all 22 Indian languages in its own evaluation.
On the IndicContextEval benchmark, Saaras V4 recorded a 16.03% WER in the L5 keyword-prompting setting, according to Sarvam and NDTV Profit.
Keyterm Prompting Helps With Difficult Words
Saaras V4 includes keyterm prompting, a feature designed to improve recognition of words that may otherwise be difficult for an ASR model to identify.
Developers can provide domain-specific terms such as:
- Company names
- Product names
- People’s names
- Acronyms
- Technical terminology
- Industry-specific vocabulary
Sarvam’s documentation says Saaras V4 can accept up to 50 keyterms through its REST and batch APIs.
This could be particularly useful in fields such as healthcare, finance, customer support and enterprise software, where correctly recognizing specific names and terminology is important.
Low-Latency Streaming for Real-Time Applications
Saaras V4 is not limited to uploading completed recordings.
Sarvam says the model supports real-time streaming, with time to first token below 150 milliseconds.
Low latency matters for applications such as voice assistants and conversational AI systems because the speech-recognition layer needs to process audio quickly enough for the rest of the application to respond naturally.
The model also supports long-form audio. Sarvam says multi-minute recordings can be processed within a second in its reported setup, although actual processing performance will depend on the application and infrastructure.
Developers Can Access Saaras V4 Through an API
Sarvam has made Saaras V4 available to developers through its speech-to-text API.
The platform supports several integration options, including:
- REST API
- Batch API
- WebSocket streaming
- Python SDK
- Node.js SDK
Sarvam also lists integrations with platforms such as Vercel AI SDK, LiveKit Agents and Pipecat Agents.
The REST API is designed for shorter audio requests, while the batch API can process files of up to two hours. WebSocket support is intended for real-time streaming applications.
This makes Saaras V4 relevant not just as a research model but as infrastructure that developers can integrate into voice applications.
Why Saaras V4 Matters for Indian AI
India’s language diversity presents a different challenge from building speech systems primarily around English.
People often switch between languages during conversations, use regional accents and dialects, and communicate in environments with significant background noise. Text may also need to be displayed in native scripts, Romanized form or English depending on the application.
Saaras V4 attempts to address several of these requirements within a single model.
Its combination of automatic language detection, 22 Indian languages, code-mixing support, transliteration, translation and real-time streaming could reduce the number of separate components developers need to assemble when building multilingual voice applications.
Saaras V4 vs Traditional Speech-to-Text Systems
A key distinction is the number of tasks Saaras V4 handles within the same model.
| Feature | Saaras V4 |
| Indian language support | 22 languages |
| English | Indian + Global English |
| Language detection | Yes |
| Transcription | Yes |
| Verbatim output | Yes |
| Transliteration | Yes |
| Translation | Yes |
| Code-mixed output | Yes |
| Noisy audio | Supported |
| Keyterm prompting | Yes |
| Streaming | Yes |
| Speaker diarization | Available through batch API |
| REST API | Yes |
| Batch processing | Yes |
| WebSocket | Yes |
This combination makes the model particularly focused on multilingual and production-oriented voice applications.
Final Takeaway
Sarvam AI’s Saaras V4 is a significant update to the company’s speech-recognition platform, with its strongest focus on multilingual Indian speech and real-world audio.
The model supports 22 Indian languages plus English, automatic language identification, five output formats, code-mixed speech, transliteration, translation, noisy audio and real-time streaming. Sarvam has also added keyterm prompting and developer APIs aimed at integrating the model into production voice applications.
For developers building voice assistants, customer-service systems, call analytics tools or multilingual applications, the ability to process different Indian languages through a single ASR model could be particularly useful.
Saaras V4 is available through Sarvam’s API, giving developers a way to test and integrate the model into their own applications.
FAQ
What is Sarvam AI Saaras V4?
Saaras V4 is an automatic speech recognition model from Sarvam AI designed to convert speech into text across Indian languages and English while handling noisy audio, code-mixing and different accents.
How many languages does Saaras V4 support?
Saaras V4 supports 22 Indian languages plus English. It also adds Global English recognition alongside Indian English.
What output formats does Saaras V4 support?
Saaras V4 supports five output modes: transcribe, translate, verbatim, translit and codemix.
Can Saaras V4 detect languages automatically?
Yes. Saaras V4 can automatically identify the spoken language and return the transcription in the corresponding native script.
Can Saaras V4 handle noisy audio?
Yes. Sarvam designed Saaras V4 to handle noisy and distorted audio, code-mixed speech and dialectal variation.
What is the Saaras V4 language identification accuracy?
Sarvam reports a language identification error rate of 5.22% across 22 Indian languages and 2.9% across the 10 most widely spoken Indian languages on its reported evaluation data.
Can developers use Saaras V4?
Yes. Developers can access Saaras V4 through Sarvam’s APIs, including REST, batch and WebSocket streaming, with Python and Node.js SDKs also available.
Source: Sarvam
