Sarvam Launches Bulbul V4 Flash, a Faster and Cheaper Voice Model for 11 Indian Languages

Sarvam AI has released Bulbul V4 Flash, a text-to-speech (TTS) model built for real-time voice applications in Indian languages. The Bengaluru company launched it on 13 October 2026 and says it starts speaking in under 250 milliseconds, runs up to 40% faster than its predecessor and costs a third less.

For anyone building a voice agent—a bank's helpline bot, a government information line or a customer-support assistant—those numbers matter. Speech that starts quickly and sounds natural is what separates a usable phone assistant from an awkward one.

What Sarvam launched

According to Sarvam and launch coverage, Bulbul V4 Flash:

• Covers 11 languages: Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and Indian English.
• Ships with 83 voices and 224 voice-and-style combinations.
• Reaches time-to-first-audio under 250 ms in production, by Sarvam's measurement.
• Is up to 40% faster than Bulbul V3 under comparable load.
• Costs ₹20 per 10,000 characters, compared with ₹30 for V3. A launch offer halves that to ₹10 until 12 November 2026.

The model is available through Sarvam's API, its Playground and Content Studio. Existing Bulbul users can switch by changing the model name in their requests; reports say V3 remains the default, so current integrations keep working.

Why latency is the headline number

In a phone conversation, people notice even short silences. A voice agent has to do three things in turn: turn the caller's speech into text, decide what to say, and turn that answer back into speech. Every step adds delay. If the speech step alone takes close to a second, the conversation feels broken.

A time-to-first-audio figure under 250 milliseconds means the system can begin speaking almost as soon as the first words of a reply are ready. It can then stream the rest while the caller listens. That leaves more of the time budget for the language model that decides what to say.

How it was built

Sarvam says Flash was distilled from a larger Bulbul V4 research model that the company pre-trained from scratch with a focus on Indian languages. Distillation trains a smaller "student" model to copy the behaviour of a bigger "teacher". The student is cheaper and faster to run, while keeping much of the teacher's quality.

That trade-off is the point of a "Flash" model. Voice agents have to respond fast, and they often handle thousands of calls at once. A smaller model cuts both delay and serving cost, which makes large deployments more affordable.

The speed and price figures are Sarvam's own. No independent benchmarks had been published at the time of writing.

Why it matters

Most of India does not interact with technology mainly in English, and many people prefer to speak rather than type. Speech technology—recognition on the way in, synthesis on the way out—is therefore central to bringing AI services to a wider population. Natural-sounding Indian-language voices have historically been hard to get right. Languages differ in script, sound and rhythm, people often mix languages mid-sentence, and good training data is scarce for some languages.

Bulbul V4 Flash is the latest part of a broader Sarvam push on voice. In September the company released Saaras V4, a speech-recognition model, along with a document-vision model. Together, the recognition and synthesis models make up the "ears" and "voice" of an Indian-language assistant, with a language model as the "brain" in between. Sarvam is one of the companies backed under the government's IndiaAI Mission to build home-grown foundation models.

Lower prices also matter for adoption. Cutting the cost of generated speech by a third makes it cheaper for start-ups, enterprises and public agencies to run voice services in regional languages at scale. That can help bring services like banking, healthcare information and grievance lines to people who are more comfortable speaking Odia, Punjabi or Malayalam than typing in English.

What to watch

The real test will be how the model performs in live deployments: whether callers find the voices natural, how well it handles mixed-language speech, and whether its latency holds up under heavy traffic. It will also face competition from global providers that are adding Indian languages to their own speech products.

Sources

• https://www.sarvam.ai/blogs/bulbul-v4-flash
• https://analyticsindiamag.com/ai-news/sarvam-ai-launches-bulbul-v4-flash-for-real-time-speech-in-11-indian-languages
• https://www.analyticsinsight.net/artificial-intelligence/sarvam-ai-bulbul-v4-flash-real-time-voice-model-for-indian-languages
• https://www.ciol.com/news/sarvam-ai-launches-bulbul-v4-flash-up-to-40-faster-33-cheaper-12674269
• https://www.newsbytesapp.com/news/science/sarvam-ai-launches-bulbul-v4-flash-faster-cheaper-text-to-speech-model/story
• https://alphasignal.ai/news/sarvam-ships-bulbul-v4-flash-a-40-faster-voice-ai-for-indian-languages