Sarvam AI, a Bangalore-based startup, has outperformed major AI players, including OpenAI and ElevenLabs, in Indian language voice recognition. The company's Saaras V3 model achieved a word error rate of approximately 19.3% on the IndicVoices top-10 language benchmarks. This surpasses the performance of OpenAI's GPT-4o Transcribe and ElevenLabs' Scribe v2, among others.
Key Factors
The Saaras V3 model was trained on over one million hours of multilingual Indian audio, with a focus on noisy speech environments and code-mixed conversations. This training data has enabled the model to handle the complexities of India's 22 scheduled languages, which span multiple script families, phonetic systems, and grammatical structures.
Performance and Applications
The performance gap between Saaras V3 and its global competitors widens further on lower-resource Indian languages, indicating its architecture and data pipeline are well-suited for this challenge. The model also supports native real-time streaming, speaker diarization, and automatic language detection, making it suitable for call centers, media companies, and enterprise customer service operations.
Broader Technology Stack
Saaras V3 is part of Sarvam's broader India-first technology stack, which includes a text-to-speech system, translation services, and a conversational agent platform. The company is one of twelve startups collaborating with the Indian government under the IndiaAI mission, a national initiative aimed at developing indigenous multilingual and multimodal AI technologies.
