Real-time voice AI in regional languages is getting faster and cheaper. Indian AI lab Sarvam AI launched Bulbul V4 Flash on October 13, 2026, a text-to-speech model covering 11 Indian languages with under 250 milliseconds to first audio in production.
According to Sarvam, V4 Flash runs up to 40 percent faster than Bulbul V3 under similar load, and pricing dropped by a third: Rs. 20 per 10,000 characters, down from Rs. 30. A launch offer halves that to Rs. 10 until November 12, 2026, so a team voicing a million characters a month would pay about Rs. 1,000. The model ships with 83 voices and 224 voice-and-style options, and Sarvam says it grew its speech data to nearly seven times the V3 amount.
The model covers Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, and Indian English, and handles mixed-language speech like Hinglish. Rather than distilling a big model, Sarvam says it trained V4 from scratch with Indian languages as the main goal, adding a text normalization layer that reads currency amounts, acronyms, and IDs the way people say them.
Developers get started through the Sarvam dashboard using the model ID bulbul:v4-flash, choosing REST, HTTP streaming, or WebSocket depending on how live the use case is. Integrations cover LiveKit, Pipecat, Twilio, Exotel, and Vapi, and Sarvam publishes migration guides for teams moving from ElevenLabs, Cartesia, Deepgram, or Gemini.
Voice builders comparing options can also look at TTSMaker, Listnr, NaturalReader, and MicMonster in our directory.
Source: Analytics Insight
