Sarvam AI's Saaras V3 Beats GPT-4o on Indian Languages

Bengaluru-based Sarvam AI has unveiled Saaras V3, a speech recognition model that has outperformed major global competitors including OpenAI's GPT-4o Transcribe, ElevenLabs' Scribe v2, Google's Gemini 3 Pro, and Deepgram's Nova-3 on Indian language benchmarks. The model achieved a word error rate of roughly 19.3% on the *IndicVoices* top-10 language tests, a notable feat given India's linguistic complexity.

India's 22 scheduled languages, multiple script systems, and widespread code-mixing (where speakers blend languages like Hindi and English mid-sentence) create conditions that typically challenge Western-trained AI models. Sarvam addressed this by training Saaras V3 on more than one million hours of multilingual Indian audio, with particular focus on noisy environments and code-mixed speech.

The model also topped the *Svarah* benchmark for Indian-accented English and supports real-time streaming, speaker diarization, and automatic language detection, making it suited for call centers and enterprise use. Saaras V3 fits into Sarvam's broader stack, which includes text-to-speech and translation tools, and the company is part of the government-backed *IndiaAI mission*.