Sarvam AI has officially unveiled Saaras V4, the latest generation of its speech recognition model. This advanced system now supports all 22 scheduled Indian languages, alongside comprehensive global English accents. The company reports that Saaras V4 achieves state-of-the-art accuracy across this diverse linguistic range.
Deployment and Technical Foundations
Saaras V4 is currently available through Sarvam AI’s API, identifiable by model="saaras:v4". While the model's internal weights are proprietary, switching from previous versions involves a simple code change. The underlying architecture features an encoder-decoder system, where an audio encoder transforms waveforms into rich phonetic and acoustic embeddings. A temporal-downsampling adapter then optimizes these sequences for the decoder.
The core of the system is Sarvam-3B, a 3-billion-parameter hybrid state-space language model developed entirely in-house. This decoder processes audio features and text prompts, generating transcripts autoregressively by feeding each token back into the system for subsequent prediction.
Benchmarking Performance
Sarvam AI has released benchmark results showcasing Saaras V4's capabilities. In English, evaluations across seven datasets, including components of Hugging Face’s Open ASR Leaderboard and AI4Bharat’s Svarah, positioned Saaras V4 with the lowest average Word Error Rate (WER) among the models benchmarked by Sarvam.
For Indic languages, tests on the Vistaar dataset utilized both WER and LLM-WER, with the latter providing semantic validation to differentiate genuine errors from formatting variations. On the Kathbath Noisy dataset, which features challenging audio conditions, Saaras V4 reportedly achieved an error rate less than half that of Deepgram Nova-3 and GPT-4o Transcribe, measured by LLM-WER. Language identification on IndicVoices utterances yielded a 2.9% error rate for the top 10 Indian languages and 5.22% for all 22. It is important to note these figures are vendor-reported and await independent verification.
Versatile Output and Keyterm Prompting
Saaras V4 offers five distinct output modes from a single model instance, streamlining post-processing workflows:
- Transcribe: Native script with normalized numbers and dates.
- Verbatim: Exact spoken words, including fillers.
- Codemix: Native script, retaining English words in Latin.
- Translit: Full utterance in Latin script.
- Translate: English translation with normalized numbers.
A new feature, keyterm prompting, allows users to provide up to 50 terms to bias the recognition engine. While not a guarantee, this enhances accuracy for specific vocabulary. For example, using codemix mode with keyterms helps maintain brand names in Latin script. On the IndicContextEval L5 benchmark, Saaras V4 achieved a 16.03% WER, reported as the lowest score on record.
Access Options and Pricing
Saaras V4 supports various interaction methods, including WebSocket for real-time streaming with a time to first token (TTFT) under 150 milliseconds, REST for synchronous transcription of clips up to 30 seconds, and Batch API for asynchronous jobs up to two hours with optional speaker diarization. SDKs are available for Python 3.9+ and Node.js 18+, alongside integrations with platforms like LiveKit Agents. Pricing is set at ₹30 per hour for standard speech-to-text and ₹45 per hour with speaker diarization, positioning it as a competitive solution in the market. Upgrading from Saaras v3 is straightforward due to identical request shapes.
Competitive Edge
Sarvam AI positions Saaras V4 as a leader, particularly in multilingual support. Its comprehensive coverage of all 22 Indian scheduled languages, alongside global English, surpasses several competitors. The model's integrated output modes and keyterm prompting offer enhanced flexibility. While some rival services provide similar features, Saaras V4's capabilities and pricing structure, at ₹30 per hour, present a compelling alternative to solutions from providers like Deepgram, ElevenLabs, and OpenAI.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
