Bangalore-based Sarvam AI outperforms OpenAI and ElevenLabs on Indian language voice recognition

A Bengaluru startup just quietly outscored some of the biggest names in AI on a task that matters enormously in a country with 22 official languages and roughly 1.4 billion people: understanding what they’re actually saying. Sarvam AI’s latest speech recognition model, Saaras V3, posted a word error rate of approximately 19.3% on the IndicVoices top-10 language benchmarks, beating OpenAI’s GPT-4o Transcribe, ElevenLabs’ Scribe v2, Google’s Gemini 3 Pro, and Deepgram’s Nova-3. In speech recognition, a lower word error rate means fewer mistakes. Why Indian languages are an AI stress test India’s 22 scheduled languages span multiple script families, phonetic systems, and grammatical structures. Add in the widespread practice of code-mixing, where speakers blend Hindi and English mid-sentence, or Tamil and English, or any number of combinations, and you get audio data that would make most Western-trained models break out in digital hives. Sarvam trained Saaras V3 on over one million hours of multilingual Indian audio, with specific attention to noisy speech environments and code-mixed conversations. The performance gap between Saaras V3 and its global competitors widened further on low...