Sarvam AI Releases Saaras V4 Speech-to-Text Model
Sarvam AI has released Saaras V4, a speech-to-text model supporting 22 Indian languages and global English, featuring keyterm prompting and achieving a 16.03% WER on September 26, 2026.
Stock photo for illustration only, not from the actual event
- Sarvam AI launched Saaras V4 supporting 22 Indian languages and global English.
- Utilizes an encoder-decoder architecture paired with the 3B-parameter Sarvam-3B language model.
- Introduces keyterm prompting supporting up to 50 terms to bias recognition accuracy.
- Achieved a 16.03% Word Error Rate on the IndicContextEval benchmark.
Sarvam AI has officially announced the release of Saaras V4, its advanced speech-to-text model designed to handle all 22 official Indian languages alongside global English, aiming to set a new benchmark for multilingual speech recognition accuracy.
Architecturally, Saaras V4 operates as an encoder-decoder system. An audio encoder translates input waveforms into embeddings carrying phonetic and acoustic details, while a temporal-downsampling adapter shortens the sequence and projects it into the language model's embedding space, enabling long recordings to fit comfortably within the decoder's context budget.
Stock photo for illustration only, not from the actual event
The decoding task is handled by Sarvam-3B, a 3B-parameter hybrid state-space language model built from scratch in-house. It processes audio features alongside text prompts and generates transcripts autoregressively by feeding each generated token back as the subsequent input.
A notable addition in V4 is keyterm prompting, which is exclusive to saaras:v4. Users can pass a JSON list containing up to 50 terms, each up to 64 characters long, to bias recognition outcomes. Additionally, the codemix mode ensures that brand names such as PhonePe remain correctly formatted in Latin script.
Building speech recognition systems capable of accurately processing dozens of linguistically diverse regional languages presents significant technical hurdles. Integrating specialized mechanisms like keyterm prompting and state-space language models represents a crucial step in mitigating transcription errors common in domain-specific and multilingual applications.
The model is currently deployable via Sarvam's API using model="saaras:v4". Developers should note that all performance figures are vendor-reported, as independent reproduction studies have not yet been published as of September 26, 2026.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment