From the source
Sarvam AI released Bulbul V4 Flash, a text-to-speech model trained on over 400 billion audio tokens and distilled for speed and cost.
The model achieves under 250 ms time-to-first-audio in production and is priced at ₹20 per 10,000 characters, a third less than the prior V3 model, while offering improved prosody and expressiveness across Indian languages.
V4 Flash was pretrained from scratch with a language-balanced curriculum to preserve native accents and speech patterns, including Indian English.
