Sarvam AI’s Bulbul V4 Flash is a text-to-speech model aimed at voice agents that need quick responses and Indian-language voices. Sarvam says it produces first audio in under 250 ms in production and costs ₹20 per 10,000 characters, one-third less than Bulbul V3. It is available through Sarvam’s API, Playground and Content Studio. Existing integrations can use the same text-to-speech endpoint, but developers must select V4 Flash and use its voice IDs.
Sarvam announced Bulbul V4 Flash on October 13. We have not tested it in a live call flow, so the latency and quality figures here are Sarvam’s.
What Bulbul V4 Flash does
Bulbul turns text into speech. V4 Flash is a smaller model distilled from Sarvam’s larger V4 research models, tuned to keep the quality while staying fast and cheap enough for live calls. Sarvam says it trained on more than 400 billion audio tokens and grew its own speech data almost seven times compared with V3.
Sarvam’s demos show the idea with a railway helpline that reads a PNR, train number, coach and berth, plus an EMI reminder that moves between Hindi and English. Its contextual text-normalisation system decides whether terms such as ISRO, EMI, PAN and Aadhaar should be spoken as a word, as letters or by their familiar pronunciation.
Sarvam also says Indian English was kept as its own accent through training, so English output should sound like an Indian speaker rather than a US voice reading Indian names.
Bulbul V4 Flash languages and voices
Sarvam lists 11 languages: Hindi, Bengali, Tamil, Telugu, Gujarati, Kannada, Malayalam, Marathi, Punjabi, Odia and English. That is 10 Indian languages plus English, not 11 Indian languages. Hinglish is supported as code-mixed text, and the voice catalogue has a separate English-Hindi group for it.
The headline figure of 11 languages needs a little unpacking. Sarvam’s model page lists V4 Flash voices for Hindi, English, English-Hindi, Bengali, Gujarati, Kannada, Marathi, Punjabi, Tamil and Telugu. Malayalam and Odia appear among the 11 supported language codes, but not in that voice list. If you are building for Kerala or Odisha, check the voice catalogue before you plan a migration.
Sarvam’s launch blog lists 83 voices and 224 voice-and-style combinations. Its API documentation describes the catalogue as 200+ voice personas and currently lists 222 voice IDs, so check the catalogue for the exact options available to your app.
Bulbul V4 Flash vs Bulbul V3
| Bulbul V4 Flash | Bulbul V3 | |
|---|---|---|
| API model name | bulbul:v4-flash | bulbul:v3 (still the default) |
| Price, pay as you go | ₹2 per 1,000 characters | ₹3 per 1,000 characters |
| Voices | 83 voices, 224 voice-style combinations | 30+ named speakers |
| Speed | Under 250 ms to first audio, up to 40% faster than V3 under comparable load (Sarvam’s figures) | Slower |
| Characters per REST request | 2,500 | 2,500 |
| Pitch and loudness controls | No (pace 0.5x to 2x) | No (pace 0.5x to 2x) |
The latency and quality claims are Sarvam’s own numbers. The company has not published a third-party benchmark, and “under 250 ms” will depend on your network path and how you stream the audio.
What ₹20 per 10,000 characters gets you
On pay-as-you-go pricing, V4 Flash costs ₹2 per 1,000 characters, the same for regular, streaming and WebSocket requests. Higher tiers bring it down to ₹1.85 on Pro and ₹1.70 on Growth and Business. Some rough examples:
- A 300-character bot reply, about two or three spoken sentences, costs around 60 paise.
- A call where the bot speaks 2,000 characters in total costs about ₹4 in speech.
- One million characters a month costs ₹2,000, compared with ₹3,000 on V3.
That is only the voice. A complete voice agent also pays for speech-to-text (₹0.49 a minute on pay as you go) and an LLM. Sarvam sells a managed Voice Agents product with everything bundled, priced from ₹3 a minute.
Sarvam’s pricing page says new accounts start with free credits, and that credits work across its APIs and never expire. It does not say how much the free credit is worth, so check your dashboard after you sign up.
Who should look at it
Bulbul V4 Flash makes the most sense for Indian businesses running voice bots, IVR menus, payment reminders, delivery updates or support lines in Hindi and regional languages. Those calls are full of names, numbers and Hinglish, which is exactly what Sarvam says it worked on. Ed-tech apps and audio content teams can use it for narration through Content Studio without writing code.
For an English-only app aimed outside India, the focus on Indian voices may not matter much. For a voice app serving Indian customers, test the languages and pronunciations you actually need before choosing a model.
How to try Bulbul V4 Flash
The quickest test is the Sarvam Playground in your browser, where you can listen to the voices without writing code. Developers get an API key from the Sarvam dashboard and call the same text-to-speech endpoint used for V3. In the Python SDK, it looks like this:
from sarvamai import SarvamAI
client = SarvamAI(api_subscription_key="YOUR_API_KEY")
response = client.text_to_speech.convert(
text="नमस्ते! आपकी EMI 5 अक्टूबर को due है।",
language_code="hi-IN",
model="bulbul:v4-flash",
)Two things trip people up. If you leave out the model name, the API still uses V3. And the old V3 speaker names such as “shubh” do not work on V4 Flash. Leave the speaker out to get the default voice, or pick a voice ID from Sarvam’s V4 catalogue. The audio comes back as base64, so decode it before you play it.
Sarvam also previewed Bulbul V4 Pro, which can add emotions such as laughter, sighs and whispers. It is not available yet, and Sarvam has not announced a price or release date. V4 Flash is the option available now for real-time use.
My take
If your app already uses Bulbul V3, V4 Flash is worth testing. Its pay-as-you-go price is lower, and Sarvam says speech quality has improved. Before switching live calls, try your own names, amounts and booking references, and confirm that the voices you need are available. The missing Malayalam and Odia voice options may matter more than the latency claim for some teams.
- SourceSarvam API pricing
