Indian artificial intelligence startup Sarvam AI has launched Bulbul V4 Flash, a text-to-speech model designed to generate natural-sounding speech quickly for real-time applications. The model supports 11 languages, including 10 Indian languages and Indian English, and delivers a reported time-to-first-audio latency of under 250 milliseconds. Sarvam has priced the model at ₹20 per 10,000 characters, compared with ₹30 for its predecessor, Bulbul V3.
The launch targets businesses developing voice agents, automated customer support, interactive voice response systems, educational tools and audio content. Sarvam says the new model improves pronunciation, pacing and contextual delivery while reducing inference costs and response delays. The company has made Bulbul V4 Flash available through its API, Playground and Content Studio, allowing developers and businesses to test or integrate the technology into their applications.
Key Takeaways
- New model: Sarvam AI has launched Bulbul V4 Flash for fast, real-time text-to-speech generation.
- Language support: 11 languages in total, including 10 Indian languages and Indian English.
- Latency: Sarvam reports time-to-first-audio below 250 milliseconds in production.
- Pricing: ₹20 per 10,000 characters, down from ₹30 for Bulbul V3.
- Performance: The company says the model can be up to 40% faster than Bulbul V3 under comparable workloads.
- Voice options: The launch announcement lists 83 voices and 224 voice-and-style combinations.
- Availability: Developers can access the model through Sarvam’s API, Playground and Content Studio.
- Business focus: The model is designed for applications where fast responses and natural speech are important, including customer service and conversational AI.
What Is Sarvam AI’s Bulbul V4 Flash?
Bulbul V4 Flash is a text-to-speech (TTS) model. It converts written text into spoken audio, allowing software to read messages, answer questions and deliver information through a voice interface.
Unlike conventional audio-generation tools designed primarily for longer recordings, Bulbul V4 Flash is built for interactive applications where a user may be waiting for the next spoken response. In a voice conversation, even a short pause can make an AI assistant feel less responsive. Reducing the time needed to produce the first audio can help conversations feel more natural.
Sarvam says it developed a smaller, more efficient model using lessons from its broader V4 research programme. The company experimented with larger language backbones, different audio representations and speech-generation approaches before applying those findings to a model optimised for production use.
The result is intended to balance speech quality, inference speed and operating cost. That balance matters for businesses deploying voice agents at scale, where a small improvement in the cost of each interaction can become significant across thousands or millions of conversations.
The model is available through Sarvam’s existing developer ecosystem, making it possible for businesses already using the company’s speech tools to explore the newer version without building an entirely separate integration.
Bulbul V4 Flash Pricing: 33% Cheaper Than Bulbul V3
One of the most important aspects of the launch is its pricing.
| Feature | Bulbul V3 | Bulbul V4 Flash |
|---|---|---|
| Price per 10,000 characters | ₹30 | ₹20 |
| Reduction in price | — | Approximately 33.3% |
| Reported time-to-first-audio | Varies by workload | Under 250 ms |
| Primary positioning | General text-to-speech | Low-latency, interactive speech |
| Language support | 11 language codes | 11 languages |
| Availability | Sarvam API | API, Playground and Content Studio |
The pricing figures are based on Sarvam’s launch announcement. Actual costs for a particular application will depend on usage, request patterns and any other services required by the deployment.
At ₹20 per 10,000 characters, generating one million characters of text would cost approximately ₹2,000 at the stated rate, compared with ₹3,000 at ₹30 per 10,000 characters. That represents a saving of about ₹1,000 for every million characters processed, assuming the same pricing applies throughout.
| Text volume | At ₹30 per 10,000 characters | At ₹20 per 10,000 characters | Indicative saving |
|---|---|---|---|
| 100,000 characters | ₹300 | ₹200 | ₹100 |
| 1 million characters | ₹3,000 | ₹2,000 | ₹1,000 |
| 10 million characters | ₹30,000 | ₹20,000 | ₹10,000 |
These are simple calculations based on the announced rates, not estimates of a complete voice-agent deployment. Businesses must also account for the costs of language-model responses, telephony, speech recognition, infrastructure and other components of a voice application.
Even so, a lower text-to-speech rate could make frequent spoken interactions more economical for companies serving large numbers of customers.
Under 250 Milliseconds: Why Low Latency Matters
Sarvam reports that Bulbul V4 Flash delivers time-to-first-audio below 250 milliseconds in production.
Time-to-first-audio measures how quickly a system begins returning generated speech after receiving the text input. It is not the same as the time required to generate an entire recording or complete a full voice-agent interaction.
The distinction matters because conversational applications often work as a pipeline. A user speaks, a speech-recognition system converts the audio into text, a language model generates a response and a TTS model converts that response back into speech.
Delays can accumulate at each stage. Faster speech generation can reduce one part of the total waiting time, but it does not guarantee that the complete system will respond within 250 milliseconds.
Sarvam says Bulbul V4 Flash uses a streaming-first inference system designed to prioritise the next section of audio rather than processing speech far ahead of the conversation. The company has also optimised data handling and reduced unnecessary synchronisation to improve performance when multiple requests are processed concurrently.
This approach is particularly relevant for voice agents that must respond quickly while handling many simultaneous conversations.
Potential applications include automated customer support, interactive voice response systems, spoken product assistance, education and conversational interfaces for digital services. In these settings, fast audio generation can improve the perceived responsiveness of an application.
The reported latency is a company-provided performance claim. Independent testing across different languages, hardware configurations and production workloads would help establish how consistently the model achieves that result.
Eleven Languages and More Natural Indian Speech
Bulbul V4 Flash supports 11 languages, including Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and Indian English.
For a country with substantial linguistic diversity, speech generation in regional languages can help make digital services accessible to people who prefer speaking or listening rather than reading long blocks of text.
Sarvam says the new model improves pronunciation, sentence pacing, contextual delivery and the expression of emotion. The company has also expanded its proprietary speech datasets nearly sevenfold compared with Bulbul V3, with broader coverage of languages, regional accents, speakers and conversational styles.
The company says its V4 research used more than 400 billion audio tokens across experiments involving different model sizes and audio-generation approaches. It used those experiments to develop a smaller production model intended to retain important speech-quality improvements while reducing latency and cost.
Context matters beyond pronunciation
Text-to-speech systems need to do more than pronounce individual words correctly. They must also interpret how names, dates, amounts, abbreviations and other written expressions should be spoken.
Sarvam says Bulbul V4 Flash uses a contextual text-normalisation layer to determine how such content should be read aloud. For example, some abbreviations should be pronounced as words, while others are better spoken letter by letter.
That distinction can be important in banking, e-commerce, travel and customer-service applications. A voice assistant that mispronounces a person’s name, booking reference or payment amount can confuse users even when the underlying information is correct.
Language support alone does not guarantee equal quality across all dialects or contexts. Businesses should test representative examples in the specific languages and use cases they intend to support.
How Businesses Could Use Bulbul V4 Flash
Customer support and voice agents
Companies can use text-to-speech to make automated support systems speak responses rather than rely exclusively on menus or text. Lower latency could help make these interactions feel more conversational, particularly when paired with fast speech recognition and language-model inference.
Banking and financial services
Financial institutions may use voice systems to communicate account information, explain processes and guide customers through routine service requests. Clear pronunciation of numbers, dates and names is especially important in these applications.
Any deployment handling sensitive financial information would also require appropriate security, privacy and compliance controls.
E-commerce and delivery services
Online retailers and delivery companies could use voice agents to answer questions about orders, provide delivery updates and assist customers who prefer speaking in a regional language.
Education and audio content
Text-to-speech can turn written material into spoken lessons, explanations and narration. Lower generation costs could make it more practical to produce larger volumes of audio, although quality, pronunciation and editorial review remain important.
Sales and multilingual business communication
Businesses may use generated speech for product explanations, appointment reminders and other routine interactions. Regional-language voices could help companies communicate with customers across different parts of India without creating every recording manually.
These are potential use cases rather than evidence that every application will achieve the same quality or cost savings. Performance will depend on the language, content, integration and surrounding application.
Bulbul V4 Flash vs Bulbul V4 Pro
Sarvam has also previewed Bulbul V4 Pro, another model in the V4 family aimed at more expressive speech generation.
The company demonstrated the ability to convey emotions such as happiness, sadness, fear and anger, alongside vocal behaviours including laughing, sighing, whispering and emphasising selected words.
Bulbul V4 Flash, by contrast, is positioned around low latency and cost-efficient generation for real-time workloads.
Sarvam has not announced a release date or pricing for Bulbul V4 Pro in the launch information reviewed. Its preview should therefore not be treated as a generally available product with confirmed commercial terms.
The two models illustrate different priorities in speech AI. A real-time customer-service assistant may prioritise fast, predictable responses, while a narration or entertainment application may place greater emphasis on expressive delivery.
For developers, the practical choice will depend on whether latency, price, voice quality or emotional range matters most to the intended application.
What the Launch Means for India’s AI Ecosystem
Sarvam’s release reflects a broader effort by Indian AI companies to build systems tailored to local languages and business needs.
Many global AI tools provide strong English-language capabilities, but deploying voice technology across India requires attention to regional pronunciation, language mixing, local names and everyday expressions. Systems that perform well in one language or accent may not automatically deliver the same results in another.
The combination of lower pricing and low-latency speech generation could make it easier for Indian businesses to experiment with voice interfaces. However, broader adoption will depend on reliability, consistent language quality, developer experience and the total cost of running complete applications.
Competition will also come from international speech providers and other Indian-language AI developers. Businesses will need to compare actual output quality, language coverage, latency, availability and integration costs rather than choosing a model solely on its advertised speed or price.
Sarvam’s reported improvements are a step in that direction, but independent evaluations and real-world deployments will be important in establishing how the model performs across different industries.
The Bigger Picture
Speech is becoming an important interface for AI because it can make digital services easier to use for people who are less comfortable with typing or navigating complex applications. For India, multilingual voice systems could help extend access to services across a wider range of languages and contexts.
The challenge is to make those systems not only fast and affordable but also accurate, natural and dependable. Bulbul V4 Flash addresses latency and pricing, while its language-focused design aims to improve the experience of spoken interactions. The significance of the launch will depend on how well those improvements translate into everyday use.
What Happens Next?
Sarvam has made Bulbul V4 Flash available through its API, Playground and Content Studio, allowing developers to test the model and integrate it into voice applications. Existing Bulbul users can migrate through the established API interface, although developers should review the current model documentation, voice identifiers and request limits before deployment.
The next important test is real-world performance across languages, accents, noisy environments and high-volume workloads. Businesses will also need to assess how the model fits into complete voice-agent systems, where speech recognition, response generation and telephony contribute to the overall user experience. Sarvam’s preview of Bulbul V4 Pro points to continued work on expressive speech, but the company has not announced its commercial release date.
Frequently Asked Questions
1. What is Sarvam AI’s Bulbul V4 Flash?
Bulbul V4 Flash is a text-to-speech AI model that converts written text into spoken audio. It is designed for real-time voice applications and supports 11 languages, including 10 Indian languages and Indian English.
2. What is the price of Bulbul V4 Flash?
Sarvam has announced pricing of ₹20 per 10,000 characters, compared with ₹30 for Bulbul V3. This represents a reduction of approximately 33.3% at the stated rates.
3. How fast is Bulbul V4 Flash?
Sarvam reports time-to-first-audio latency below 250 milliseconds in production and says the model can be up to 40% faster than Bulbul V3 under comparable workloads. These are company-reported performance figures.
4. How can developers access Bulbul V4 Flash?
The model is available through Sarvam’s API, Playground and Content Studio. Developers can consult Sarvam’s documentation for supported languages, voice options, integration methods and request limits.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
