Bulbul v3 is Sarvam AI's Indic-first TTS: 11 Indian languages, 30+ persona voices (shubh, ishita), 8–48kHz output, and ₹30 per 10,000 characters in beta — roughly 10x cheaper than Western per-character rates. Backed by the IndiaAI Mission, with no self-serve cloning and honest per-language voice-quality variance. Runs in the invideo agent's Audio tab for dubbing, regional voiceover, and EdTech narration.
Updated August 2026
Sarvam Bulbul is the text-to-speech model family from Sarvam AI, the Bengaluru company building India's sovereign AI stack, and Bulbul v3 is its current generation: 30+ persona voices across 11 Indian languages, trained — in Sarvam's own words — "on Indian languages and culturally grounded data." At ₹30 per 10,000 characters in beta (August 2026), it undercuts Western per-character TTS rates by roughly an order of magnitude, which is precisely the point: this is TTS priced and tuned for regional-language Indian content, not a global model with Hindi bolted on.
Who is Sarvam AI?
Sarvam AI was founded in Bengaluru in August 2023 by Vivek Raghavan and Pratyush Kumar, both out of the AI4Bharat research effort at IIT Madras. In April 2025 it was selected under the Indian government's IndiaAI Mission to build indigenous foundation models with subsidized GPU capacity, and in June 2026 it reportedly became a unicorn — a $234M Series B at a $1.5B valuation led by HCLTech, per Wikipedia's account of press coverage (corporate history here is secondary-sourced; treat the funding specifics accordingly). Beyond Bulbul, the company ships open-source LLMs (Sarvam-30B/105B), the Saaras v3 speech-to-text model covering 23 languages, and a vision model.
Bulbul, in other words, is not a side project — it's the speech layer of a deliberately India-first stack.
What languages and voices does Bulbul v3 support?
Bulbul v3 covers 11 languages, per Sarvam's official docs (August 2026): Indian English (en-IN), Hindi, Bengali, Tamil, Telugu, Kannada, Malayalam, Marathi, Gujarati, Punjabi, and Odia.
Voices are persona-styled rather than anonymous IDs — 30+ of them, lowercase-keyed in the API. The docs describe shubh as a "deep, dramatic tone suited to suspense or villain characters," while ishita is pitched at natural conversation. One detail straight from Sarvam's own documentation: voice quality varies by language, and the docs publish per-language speaker recommendations rather than pretending every voice performs equally everywhere. That candor is unusual and useful — check the recommended voices for your target language before committing to one.
| Spec | Bulbul v3 (Aug 2026) |
|---|---|
| Languages | 11 (en-IN + 10 Indic) |
| Voices | 30+ persona voices |
| Max input | REST ≤2,500 chars; HTTP streaming ≤3,500; WebSocket ≤2,500 (<500 recommended for lowest latency) |
| Sample rates | 8 kHz – 48 kHz |
| Formats | MP3, WAV, AAC, OPUS, FLAC, PCM, MULAW, ALAW |
| Controls | Pitch, pace, speaker selection |
| Voice cloning | Not offered (no self-serve cloning product) |
| Pricing | ₹30 per 10K characters (beta); ₹100 free credits for new users |
Read the format row closely: MULAW and ALAW are telephony codecs, and 8 kHz is a phone-line sample rate — IVR and call-center audio are first-class targets here, not an afterthought.
How much does Sarvam Bulbul cost?
₹30 per 10,000 characters under beta pricing as of August 2026, per Sarvam's official pricing page — double the ₹15 rate of the previous Bulbul v2, and still remarkably cheap: at current exchange rates that's roughly $3.60 per million characters, where prominent Western TTS APIs charge tens of dollars per million characters, an order-of-magnitude gap. New accounts get ₹100 in free credits, and rate limits scale from 60 to 1,000+ requests per minute by tier. Beta pricing can change; the direction of the offering — INR-priced, volume-friendly — is the durable fact.
Why does an India-tuned TTS model matter?
Because accent and code-mixing are where general-purpose TTS breaks. Sarvam positions Bulbul as "purpose-built" for Indian languages, "trained and evaluated on Indian languages and culturally grounded data" — which shows up in practice as native-sounding Indic pronunciation and comfort with code-mixed Hinglish, the register most real Indian scripts are actually written in. For the fastest-growing segments of Indian content — regional-language YouTube channels, EdTech courseware in mother tongues, IVR and customer-service lines, and dubbing existing content across state languages — a voice that sounds local is not a nice-to-have; it's the difference between watchable and not.
The trade-offs: 11 TTS languages is fewer than the 23 their own Saaras STT model hears, so some Indic languages can be transcribed but not yet spoken; per-language voice quality varies (their docs say so); and there is no self-serve voice cloning — you pick from the catalog. The absence of cloning also means there's no cloning-consent surface to worry about: you cannot misuse what isn't offered.
What can you make with Bulbul v3?
- Regional-language dubbing — re-voice existing videos into Hindi, Tamil, Bengali, or any of the 11 languages through an AI dubbing workflow, using Sarvam's per-language recommended voices.
- Voiceovers for regional YouTube and short-form — script in Hinglish or a regional language and generate narration in an AI voice generator flow, at a per-character cost that survives daily publishing schedules.
- EdTech and courseware — mother-tongue narration at scale for lessons built with an educational video maker, where the ₹-denominated pricing keeps per-lesson audio costs near zero.
Bulbul v3, in short
What languages does Sarvam Bulbul v3 support?
Eleven: Indian English, Hindi, Bengali, Tamil, Telugu, Kannada, Malayalam, Marathi, Gujarati, Punjabi, and Odia, per the official docs as of August 2026.
How much does Bulbul v3 cost per character?
₹30 per 10,000 characters under beta pricing (about $3.60 per million characters), with ₹100 in free credits for new users. Bulbul v2 remains listed at ₹15 per 10,000.
Can Sarvam Bulbul clone my voice?
No. Sarvam documents no self-serve voice-cloning product for Bulbul as of August 2026 — you choose from the 30+ catalog voices.
Does Bulbul handle Hinglish and code-mixed scripts?
Yes — Indic-first, culturally grounded training is Sarvam's stated positioning, and mixed Hindi-English scripts are a primary use pattern for its conversational voices.
Is every Bulbul voice equally good in every language?
No, and Sarvam says so itself: the docs note voice quality varies by language and publish per-language speaker recommendations. Use the recommended voices for your target language.
Where can you use Sarvam Bulbul v3?
Bulbul v3 is one of the speech models wired into invideo — it appears in the agent's Audio tab as "Sarvam Bulbul v3 — Indian-language TTS across 11 languages," alongside the platform's 200+ model lineup. The practical upshot for Indian creators: pick a Bulbul voice, and the narration lands straight on a video timeline with the dubbing and voiceover workflows linked above, no separate Sarvam API account required.
Version history: Bulbul v2 (₹15/10K chars, still listed) → Bulbul v3 with 30+ voices and 11 languages, beta-priced at ₹30/10K chars (current as of August 2026). Company milestones: IndiaAI Mission selection April 2025; reported $1.5B-valuation Series B June 2026 (secondary-sourced). Facts per official Sarvam AI documentation unless noted.