The new text-to-speech model delivers audio in under 250 milliseconds, cuts generation costs by one-third compared with its predecessor, and targets real-time voice applications across industries.
India’s artificial intelligence ecosystem is taking another step towards more accessible, multilingual voice technology as Bengaluru-based Sarvam AI introduces Bulbul V4 Flash, a text-to-speech model designed to generate natural-sounding speech in 11 Indian languages with exceptionally low latency. With time-to-first-audio latency under 250 milliseconds and a lower price than its predecessor, the model aims to make AI-powered voice interactions faster, more affordable, and better suited to real-world business applications.
The launch comes as businesses increasingly adopt conversational AI for customer service, automated assistance, education, banking, e-commerce, and sales. In these settings, the speed and quality of spoken responses can significantly influence how naturally people interact with AI systems.
Sarvam AI has priced Bulbul V4 Flash at ₹20 per 10,000 characters, down from ₹30 for Bulbul V3. This represents a reduction of approximately 33%, potentially lowering operating costs for companies that generate large volumes of speech through automated systems.
According to the company, the new model also improves pronunciation, speech pacing, and contextual delivery while offering performance optimisations intended for production-scale deployments.
Faster Speech Generation at a Lower Cost
Latency is a critical factor in conversational AI. When a virtual assistant takes too long to respond, the interaction can feel unnatural, interrupt the user’s workflow, or reduce customer satisfaction. By delivering its first audio in under 250 milliseconds, Bulbul V4 Flash is designed to support more responsive, real-time conversations.
Sarvam reports that the model is up to 40% faster than Bulbul V3 under comparable workloads. The company has also refined its inference infrastructure to manage multiple simultaneous conversations, an important capability for enterprises operating voice agents at scale.
Rather than generating more speech than an interaction immediately requires, the system uses a streaming-first inference approach that prioritises the next portion of audio needed for the conversation. Sarvam has also reduced unnecessary synchronisation and data copying to improve performance when handling concurrent requests.
These changes are intended to balance three important requirements for enterprise voice technology: speed, affordability, and speech quality.
The model is available through Sarvam’s API, Playground, and Content Studio. Existing Bulbul users can migrate to V4 Flash through the existing Bulbul API, helping developers adopt the updated model without having to start their integrations from scratch.
A Stronger Focus on Indian Languages and Regional Speech
A key focus of Bulbul V4 Flash is improving how AI systems speak Indian languages and represent the diversity of voices, accents, and conversational styles found across the country.
Sarvam says it has expanded its proprietary speech datasets to nearly seven times the size of those used for Bulbul V3. The expanded data covers a broader range of languages, regional accents, speakers, emotional registers, and speaking styles.
The company also pretrained the V4 model from scratch, placing particular emphasis on Indian-language representation and a language-balanced training process. The objective is to improve pronunciation while preserving the distinctive characteristics of individual languages.
This matters because multilingual text-to-speech systems must do more than convert written words into sound. They need to account for differences in pronunciation, rhythm, intonation, names, numbers, and everyday expressions. A model that handles these elements poorly may sound unnatural or produce confusing results, even when the underlying text is correct.
Sarvam says Bulbul V4 Flash is designed to make regional voices sound more natural and ensure that Indian English retains its characteristic pronunciation. It also aims to handle names, numbers, and common expressions in ways that better reflect how people actually speak.
The model offers 83 voices and 224 voice-and-style combinations, giving developers flexibility to select voices and delivery styles for different audiences and use cases.
Context-Aware Speech for Real-World Applications
One of the model’s important features is its contextual text-normalisation layer, which helps determine how written information should be spoken aloud.
Names, dates, monetary amounts, abbreviations, and reference numbers can be interpreted differently depending on their context. Some abbreviations are pronounced as complete words, while others are spoken letter by letter. Similarly, numbers may need to be rendered differently depending on whether they represent a price, a date, an account reference, or another type of information.
For AI voice agents, these distinctions are more than minor technical details. A mispronounced customer name, an incorrectly spoken payment amount, or a confusing booking reference can undermine trust and create unnecessary friction.
By improving how text is interpreted before speech generation, Bulbul V4 Flash aims to make automated conversations clearer and more reliable.
This capability could prove particularly valuable in banking, financial services, e-commerce, customer support, and other environments where voice agents communicate sensitive or transactional information. It may also support educational platforms, sales teams, and content creators that need to deliver spoken information across multiple Indian languages.
Research-Driven Development Behind Bulbul V4 Flash
Sarvam’s latest release builds on research into larger language backbones, alternative audio representations, and generative approaches to speech synthesis. The company says its research involved training data scaled to more than 400 billion audio tokens.
Rather than deploying a model that simply reflects the scale of those experiments, Sarvam used the findings to develop a smaller model intended to preserve key improvements in speech quality while meeting practical requirements for latency, cost, and production deployment.
This approach reflects a broader challenge in enterprise AI development. Larger models can offer sophisticated capabilities, but businesses also need systems that respond quickly, operate efficiently, and handle substantial volumes of requests without excessive infrastructure costs.
For text-to-speech applications, those requirements become particularly important when thousands of customers may be interacting with automated voice systems simultaneously.
Bulbul V4 Flash is positioned to address this balance by combining speech-quality improvements with an inference architecture designed for real-time use.
Bulbul V4 Pro Preview Signals a Move Towards More Expressive AI Voices
Alongside the Flash release, Sarvam has previewed Bulbul V4 Pro, another model in the V4 family focused on more expressive speech generation.
The preview demonstrated the model’s ability to convey emotions such as happiness, sadness, fear, and anger. It also showcased expressive vocal actions, including laughing, sighing, whispering, and placing emphasis on particular words.
While Bulbul V4 Flash prioritises speed, cost efficiency, and real-time deployment, the Pro preview points towards a broader ambition: developing AI voices that can communicate not only the correct words but also a more appropriate emotional tone.
Such capabilities could be relevant to interactive storytelling, digital learning, entertainment, virtual assistants, and customer-facing applications where expressive delivery can make conversations feel more engaging.
However, Sarvam has not announced a release date or pricing for Bulbul V4 Pro. Its availability and final capabilities therefore remain unconfirmed.
What Bulbul V4 Flash Could Mean for India’s Voice AI Market
India’s linguistic diversity creates a significant opportunity for AI systems that can communicate effectively beyond English. Businesses serving customers across different regions need voice technologies that can accommodate local languages and speaking patterns without making every interaction expensive or technically complex.
Bulbul V4 Flash could help developers address this requirement through a combination of multilingual coverage, a broad selection of voices, lower character-based pricing, and faster audio generation.
For startups, the lower price may make it more practical to build voice-first products. For established enterprises, improved concurrency and lower latency could support the expansion of automated customer service and other voice-based workflows. Educational platforms and content producers may also benefit from being able to generate spoken material in different languages and styles.
Nevertheless, real-world adoption will depend on factors beyond headline speed and pricing, including language-specific accuracy, performance under heavy workloads, integration requirements, and the quality of speech in different conversational contexts.
The launch also highlights an important direction for AI development in India: building models around the linguistic needs and everyday communication patterns of local users rather than treating Indian languages as secondary additions to systems primarily designed for English.
With Bulbul V4 Flash, Sarvam AI is combining faster speech generation, a lower cost per character, and a stronger emphasis on Indian-language representation. As the company develops its broader Bulbul V4 family, the next test will be how effectively these capabilities translate into dependable, natural, and scalable voice experiences for businesses and consumers.
Read more:Â Carley Mchome: Building a Family Where Every Child Has a Place to Call Home
