Introduction

On October 13, Sarvam AI introduced Bulbul V4 Flash, a closed text-to-speech model aimed at real-time voice agents across Indian languages. It is live today on the Sarvam API as bulbul:v4-flash, in the API Playground, and in Content Studio. Existing Bulbul callers switch on the same endpoint. There are no open weights — this is API-only.

Sarvam’s pitch is blunt: native-sounding Indian speech that is fast enough for live calls and cheap enough to run at volume. The launch post puts production time-to-first-audio at under 250 ms, list pricing at ₹20 per 10,000 characters (a third below Bulbul V3’s ₹30), and a launch discount that cuts that to ₹10 per 10K characters through November 12, 2026. Docs currently banner the same half-price window.

What Shipped

  • Model id: bulbul:v4-flash on the familiar Bulbul TTS API.
  • Latency: under 250 ms time-to-first-audio in production; Sarvam says Flash is up to 40% faster than V3 under comparable load.
  • Voice library: 83 voices across 11 languages, with 224 voice-and-style combinations spanning customer care, sales, collections, banking, ecommerce, education, narration and more.
  • Training scale: distilled from larger V4 research runs trained on more than 400 billion audio tokens. Proprietary speech data is about 7× larger than V3’s, with broader coverage of languages, accents, speakers and emotional registers.
  • Runnable today: Sarvam API, Playground TTS, Content Studio. No Hugging Face weights from the sarvamai org for Flash.

Built for Indian Conversations

Flash is the production cut of Sarvam’s V4 research. They trained larger backbones first, then distilled what listeners notice most into a smaller streaming model. V4 was pretrained from scratch with Indian-language representation as a primary goal, then kept on a language-balanced curriculum so languages do not blur into each other. A final per-voice accent phase tunes each speaker’s native rhythm.

That shows up in the details voice agents actually speak:

  • Context-aware text normalization for names and numbers — ISRO as a word, EMI letter by letter, plus PAN, Aadhaar, PNR and train numbers — without breaking mid-sentence Hinglish.
  • Indian English kept in balance through every training stage, not bolted on at the end.
  • Streaming-first inference meant for thousands of concurrent calls: receptive-field-aware streaming, real-time pacing, and serving-path cuts that Sarvam says never trade correctness for speed.

The launch demos lean into railway booking, EMI reminders and Bengaluru / Kolkata conversational reads. Those are product samples, not independent benchmarks, but they are the workload Flash is clearly aimed at.

Pricing and What’s Next

  • List price: ₹20 per 10,000 characters.
  • Launch price: ₹10 per 10,000 characters through November 12, 2026 (50% off), confirmed on Sarvam’s launch thread.
  • V3 reference: ₹30 per 10,000 characters.
  • Latency claim: under 250 ms time-to-first-audio in production.

Related Sarvam moves from the same week (context only, not this drop): Instant Voice Cloning landed in the Python and TypeScript SDKs on October 12, and Bulbul V4 itself reached Voice Agents on October 6. Separately, Sarvam teases Bulbul V4 Pro — a growing V4-family model that can laugh, sigh, sound nervous and stress individual words. Treat that as a preview, not a ship date.

Conclusion

If you are building Indian-language voice agents and already on Sarvam, Flash is a same-day API swap with a clear latency and price story through mid-November. If you need open weights or local serving, this is not your drop — stay on open TTS stacks and watch whether Sarvam ever mirrors Flash to Hugging Face. For everyone else, hit the Playground with a Hinglish line that mixes a PNR and an amount, and listen for whether the numbers and code-switches land the way your callers expect.

—Anabel