Introducing
Bulbul V4 Flash

Today, we're introducing Bulbul V4 Flash, our latest text to speech model that brings more natural and expressive speech across Indian languages, with the speed and efficiency needed for real-time applications.

With V4 Flash, our research focused on capturing the richness and nuances of Indian languages, while making high-quality speech generation faster and more affordable.

We've made substantial improvements in how expressive speech is generated across Indian languages. The model is better at capturing emotion, adjusting emphasis, pacing sentences and responding to the context of what is being said. We've focused on bringing greater depth and expressiveness across the breadth of Indian languages, preserving the distinct character of each language and the subtleties that make speech sound natural to its speakers.

Alongside improvements in speech quality, we've developed a smaller, more efficient model and optimised the inference stack around it. In real-time voice applications, every delay affects the flow of a conversation. Fast, predictable responses allow conversations to flow naturally, even when thousands of calls are happening at once.

Flash delivers under 250 ms time-to-first-audio in production and costs ₹20 per 10,000 characters.

Frontier quality in a smaller model

Flash is an outcome of our V4 research. We set aside the latency and cost limits of a production model to see how far speech quality could go. We trained larger language backbones, explored richer audio representations, tested different generative approaches and scaled training data to more than 400 billion audio tokens.

Quality had not plateaued. Even at that scale, more data and more capable models kept improving the model’s use of context, along with its prosody, expressiveness and robustness.

To build Flash, we looked across several experimental iterations, identified what drove the gains, and distilled it into a much smaller, more efficient model. Flash keeps the qualities listeners notice most while meeting the speed and cost requirements of production.

Bulbul V3 was priced at ₹30 per 10,000 characters. Flash costs ₹20, a third less, and it sounds better, with richer prosody, stronger contextual delivery and higher speech quality. At production volume, expressive speech no longer has to give way to cost.

Built for the breadth and depth of Indian Languages

Speech data across Indian languages varies widely in volume, speaker diversity and expressive range. We expanded our proprietary speech datasets nearly seven times relative to V3, increasing coverage across languages, regional accents, speakers, emotional registers and conversational styles.

Scale solved one problem and exposed another. A compact model trained on many languages tends to average them, blurring the features that make each one distinct. So we pretrained V4 from scratch with Indian language representation as a primary objective. We then trained it with a continuous, language-balanced curriculum designed to reinforce native pronunciation, speaker identity and expressive delivery throughout training.

As a result, regional voices sound native, Indian English sounds Indian, and names, numbers and everyday expressions come out the way people say them.

Native, down to the detail

A voice sounds native when it gets the small things right. A Bengali speaker’s "ধুর বাবা". The respect in a Telugu “garu”. Their meaning comes through pauses, emphasis and intonation, and through the ease with which speakers move between languages.

V4 learns these patterns across its multilingual curriculum, where every language is reinforced throughout training rather than added at the end. A final voice-specialization phase then tunes each voice to its native accent, conversational rhythm and expressive delivery, so that a Bengali voice carries Bengali parlance in every sentence it reads.

Indian English, as it’s spoken

For many conversations in India, English is the working language. It should sound like the English India speaks.

Indian English has its own pronunciation, stress and intonation, shaped by the regional accents that run through it. For V4, we kept Indian-English speech in balance through pretraining and every stage after it.

Context-aware by design

Much of what voice agents say is names, amounts, dates and reference numbers. The correct spoken form depends on context. ISRO is said as a word, “is-row”, while EMI is spoken letter by letter. PAN is pronounced as a word, and Aadhaar has its own familiar pronunciation.

To address this, we built a real-time contextual text normalization layer. It interprets text in context, identifies what each element represents, and determines how it should be spoken. Expressive alignment during training then ties that accuracy to the model’s delivery, so the details come through correctly without interrupting the flow of the conversation.

Listen for the PNR, train number, coach and seat, each read the way it should be, alongside PAN, Aadhaar, dates and times, all within a fluid Hindi-English conversation. When a customer’s name, a payment amount or a booking reference is spoken correctly, the conversation keeps moving. That is what language depth means in Bulbul V4 Flash. Every detail sounds right to the person on the other end of the call.

Listen for context-aware pronunciation of the PNR, train number, coach and seat, alongside natural handling of PAN, Aadhaar, dates and time, all within a fluid Hindi-English conversation.

For voice agents, this distinction matters. An incorrectly spoken customer name, payment amount or booking reference isn't just unnatural; it can make an otherwise helpful interaction confusing.

V4 brings together native pronunciation, contextual understanding and expressive speech, so that real-world information can be delivered both accurately and naturally.

That's language depth in Bulbul V4 Flash: not just speaking India's languages, but speaking them the way India does.

Low latency at scale

Bulbul V4 Flash delivers under 250 ms time-to-first-audio in production.

Flash runs on a streaming-first inference stack. Receptive-field-aware streaming skips any work that cannot affect the next stretch of audio. Real-time pacing keeps compute on what active conversations need now, instead of generating far ahead of them.

We also tuned the serving path for many simultaneous conversations. We cut unnecessary synchronization and copying, grouped compatible streaming work, and balanced requests by live load. Every optimization was gated on output correctness, so speed never came at the cost of speech quality.

Together with V4’s architectural improvements, these changes make Flash upto 40% faster than V3 under comparable load, with substantially more headroom for concurrent conversations.

One model. Many experiences.

The most exciting part of V4 Flash is seeing the same underlying model become completely different products.

And this is only the beginning.

V4 Flash launches with 83 voices across 11 languages and 224 voice-and-style combinations, spanning customer care, sales, collections, banking, ecommerce, education, conversational experiences, narration and more.

A voice for every experience.

And we are not done

Try Bulbul V4 Flash

Bulbul V4 Flash is available today through the Sarvam API, Playground and Content Studio.

• <250 ms time-to-first-audio
• ₹20 per 10,000 characters
• 11 languages
• 83 voices
• 224 voice and style combinations

Existing Bulbul users can move to V4 Flash using the familiar Bulbul API.