Bulbul V4 Flash
India's voice model for Indian languages: sub-250ms streaming TTS across 11 languages with Hinglish code-switching and native name pronunciation — ₹20 per 10K characters.
About
Bulbul is Sarvam AI's text-to-speech model family, and V4 Flash (announced September 30, 2026) is the current speed build: first audio in under 250ms, priced at ₹20 per 10,000 characters (down a third from V3's ₹30), with 83 voices and 224 voice-and-style combinations across Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, and Indian English. Sarvam says V4 was pretrained from scratch on a speech dataset roughly 7x the size of V3's, with language-balanced training aimed at the things generic TTS mangles — Indian names, currency amounts, abbreviations, and code-mixed Hinglish/Tanglish. The earlier V3 topped a Josh Talks blind listening study with 20,000+ votes (Sarvam's claim). It's API-only — no open weights — served over HTTP streaming, WebSocket, or single-response endpoints with Python/Node SDKs. For dubbing and voice work targeting Indian audiences, it's the most natively-built option in the market.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Top-ranked in a Josh Talks blind study with 20,000+ votes (provider-claimed, Bulbul V3)\n- Sub-250ms time-to-first-audio and ~40% faster than V3 (provider-claimed)\n- No independent benchmark replication found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (TIER metrics, metrics-v1). Evidence: official Sarvam docs and pricing pages; launch covered by multiple tech outlets (Oct 2026). Strengths: purpose-built for Indian languages with code-switching, transparent pricing, live API with SDKs. Docked for: (1) closed and proprietary — no weights, no independent evals; (2) quality and latency figures are all Sarvam's own; (3) Indian-languages-only scope limits generality. License: proprietary.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 1.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Bulbul V4 Flash announced 2026-09-30: 11 Indian languages, sub-250ms TTFB, Rs.20/10K chars (promo Rs.10 through 2026-11-12)
- · 83 voices, 224 voice-style combos; Hinglish/Tanglish code-switching, Indian name pronunciation
- · Blind-study win (20K+ votes) was for Bulbul V3 and is Sarvam's own claim; no independent benchmarks
- · API-only, no open weights; Indian-languages-only scope