Sarvam AI Unveils Bulbul V3: A Leap in Natural Speech Generation
Indian artificial intelligence startup Sarvam has officially launched Bulbul V3, its latest text-to-speech AI model, as part of an ambitious 14-day rollout of new AI tools. This release aims to significantly enhance the naturalness of AI-generated speech by incorporating prosodic elements such as pauses, emphasis, pacing, and tone modulation.
Advanced Features and Language Support
Bulbul V3 offers over 35 high-quality voices sourced from professional voice artists, with support for more than 11 Indian languages. The company has announced plans to extend this support to all 22 scheduled Indian languages in the near future. Built on a large language model (LLM), Bulbul V3 analyzes text and converts it into speech that mimics human-like nuances, making it ideal for applications requiring realistic audio output.
Key capabilities include:
- Low-latency streaming for real-time audio generation and playback, crucial for conversational applications and live interactions.
- Voice cloning feature with built-in safeguards, designed for consent-based, high-volume enterprise use cases.
- Handling of complex Indian speech patterns, including language switching mid-sentence, regional accents, and emotional nuances.
Testing and Performance Metrics
In a rigorous evaluation, Bulbul V3 was assessed by an independent third-party in a blind A/B human listening study across 11 languages. The study compared audio samples generated by Bulbul V3 against competitors' models using identical input text.
Performance highlights:
- Bulbul V3 outperformed Cartesia Sonic-3 and other rivals in general full-band evaluations.
- It achieved the top position in 8 kHz telephony evaluations, surpassing all other models.
- The model demonstrated the lowest rates of word skips and mispronunciations while maintaining comparable performance on extra-content errors.
Strategic Context and Future Developments
This launch is part of Sarvam's 14-day blitz leading up to the India-AI Impact Summit 2026, scheduled from February 16 to 20 in New Delhi. Sarvam is among 12 startups selected by the Indian government to develop sovereign LLMs under the Rs 10,300-crore India AI Mission, with these indigenous models expected to be unveiled at the summit.
For developers and users, Bulbul V3 is accessible via the Sarvam Dashboard, with unlimited API access available until February 28, 2026. This initiative underscores Sarvam's commitment to advancing AI technology tailored for India's diverse linguistic landscape.
Recent AI Releases from Sarvam
In addition to Bulbul V3, Sarvam has introduced several other AI tools in recent days:
- Sarvam Vision: A 3 billion-parameter vision-language model for tasks like image captioning and chart interpretation.
- Sarvam Samvaad: Conversational AI agents integrated with enterprise tools for data-driven insights.
- Sarvam Audio: An audio extension of the Sarvam 3B model, pre-trained on English and 22 Indian languages.
- Sarvam Dub: An AI dubbing model with zero-shot voice cloning for multilingual content creation.
