Sarvam launches Bulbul V4 Flash text-to-speech for Indian languages

What Sarvam released

Sarvam AI has released Bulbul V4 Flash, a text-to-speech model for Indian languages built for real-time voice agents. Sarvam’s launch post, dated 13 October 2026, says the model delivers under 250 milliseconds of time-to-first-audio in production and costs ₹20 per 10,000 characters, a third less than the ₹30 Sarvam charged for Bulbul V3. The Economic Times’ enterprise AI desk reported the launch the following morning, India time.

The model is available now through the Sarvam API, the Playground and Sarvam’s Content Studio. The company’s text-to-speech page lists eleven languages (Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and Indian English) and 83 voices in 224 voice-and-style combinations, covering uses such as customer care, sales, collections, banking, ecommerce, education and narration.

Sarvam describes itself as an Indic language AI company, and Bulbul is the speech-output half of its voice stack. The pitch for V4 Flash is aimed squarely at companies running phone agents in India: sound more natural in the languages and accents callers actually use, respond fast enough that a conversation does not stall, and cost less per call.

How Sarvam says it built V4 Flash

Sarvam says Flash came out of a broader V4 research effort in which it set aside production limits on latency and cost to see how far speech quality could go. It trained larger language backbones, tried richer audio representations and different generative approaches, and scaled training data to more than 400 billion audio tokens. According to the post, quality had not plateaued even at that scale.

Flash is the production distillation of that work. Sarvam says it looked across several experimental iterations, identified what drove the gains and distilled them into a much smaller, more efficient model. It has not published the parameter count, the architecture or the size of the larger research models.

On data, Sarvam says it expanded its proprietary speech datasets nearly seven times relative to V3, adding languages, regional accents, speakers, emotional registers and conversational styles. Because a compact model trained on many languages tends to average them together, the company says it pretrained V4 from scratch with Indian-language representation as a primary objective, then used a continuous, language-balanced curriculum and a final voice-specialisation phase that tunes each voice to its native accent and rhythm. That emphasis on balanced Indic data echoes other recent efforts, such as the 2.8-trillion-token Indic pretraining corpus NVIDIA and AI4Bharat detailed this week.

Names, numbers and code-mixed speech

Much of what a voice agent says is not prose but data: names, amounts, dates, PNR numbers and account references. Sarvam says it built a real-time contextual text-normalisation layer that works out what each element represents and how it should be spoken. Its examples are specific to India. ISRO is said as a word, EMI is spoken letter by letter, PAN is pronounced as a word and Aadhaar has its own familiar pronunciation.

The launch page demonstrates this with sample conversations rather than metrics. One is a Hindi rail-helpline agent that checks a waitlisted ticket, offers a seat on another train and books it, with short pauses where the agent waits on a tool call. Another is an EMI overdue reminder that reads out an amount, a due date and a payment cut-off. Others show a Bengali speaker complaining about Kolkata traffic and a very Bengaluru stretch of Indian English, which Sarvam says it kept in balance through pretraining because English is the working language of many Indian conversations.

Code-mixed text such as Hinglish is supported, with a caveat in Sarvam’s documentation: each language should be written in its own script. The docs warn that romanised Indic input, such as “Aapka order confirm ho gaya hai” typed in Latin letters, significantly reduces output quality.

Speed and serving

Sarvam attributes the sub-250-millisecond figure to a streaming-first inference stack. The post says receptive-field-aware streaming skips any work that cannot affect the next stretch of audio, and real-time pacing keeps compute on what active conversations need now instead of generating far ahead.

It also says it tuned the serving path for many simultaneous calls by cutting unnecessary synchronisation and copying, grouping compatible streaming work and balancing requests by live load, with every optimisation gated on output correctness. Together with the architectural changes, Sarvam claims Flash is up to 40 percent faster than V3 under comparable load, with more headroom for concurrent conversations. These are Sarvam’s own production figures; it has not published a benchmark method, hardware details or results from an independent listening test.

How developers use it

Sarvam’s Bulbul model documentation says V4 Flash uses the same request and response contract as Bulbul V3 on the same /text-to-speech endpoint, so switching is mainly a matter of setting the model to bulbul:v4-flash and choosing a speaker. Bulbul V3 remains the API default when no model is specified, which means existing integrations will not change voice unless developers opt in.

Voices are organised as personas rather than plain names. The voice catalogue says most IDs follow a name, language and style pattern, such as simran_en_customer or ishita_hi_devotional, and the default persona is shubh_enhi_ads, an English and Hindi code-mixed voice. Sarvam also points developers to a beta instant voice-cloning API that generates speech from a short reference clip.

The documented limits are practical ones. REST requests accept up to 2,500 characters and sample rates up to 48 kHz; HTTP streaming and WebSocket connections top out at 24 kHz, and output defaults to 22,050 Hz. A pace control runs from 0.5 to 2.0, but Bulbul does not support SSML tags, so fine control over pauses and emphasis means splitting text at natural break points. Sarvam also sells a managed Voice Agents platform with telephony and analytics starting at ₹3 per minute, according to its docs.

Where it fits among voice models

The release lands in a busy fortnight for speech AI. xAI opened Grok Voice 3.0 in its Voice Agent API and OpenAI began letting developers create GPT-Live voices from a text description. Those are general-purpose systems from US labs. Sarvam’s bet is narrower and more local: a price quoted in rupees, voices tuned to Indian accents and registers, and pronunciation rules for the identifiers Indian customers hear on every banking or travel call.

Language-specific behaviour is not just a voice-quality issue. A study we covered this week found that LLMs refuse more harmful requests when told they are being safety-tested in Indic languages, a reminder that models can act differently across languages in ways that English-only testing misses. Teams deploying voice agents in several Indian languages will still need to test each one with real callers.

Limitations and what to watch

The headline numbers do not line up perfectly across Sarvam’s own pages. The launch post and product page say 83 voices across 11 languages, while the developer documentation lists 222 persona voice IDs and says V4 Flash personas cover Hindi, English, English-Hindi, Bengali, Gujarati, Kannada, Marathi, Punjabi, Tamil and Telugu, and are not yet offered for every V3 language, naming Malayalam and Odia as languages V3 still covers. Developers who need Malayalam or Odia should check the catalogue before migrating.

Quality claims are also unverified. Sarvam says Flash sounds better than V3, with richer prosody and stronger contextual delivery, but it has not released mean-opinion-score results, a word-error comparison or a third-party evaluation, and the model is available only as a paid API rather than open weights. Its samples are curated demos.

Sarvam is already teasing the next step. It says a larger model in the V4 family, Bulbul V4 Pro, can laugh, sigh, sound nervous and stress a particular word, and it posted a short Hindi preview with inline emotion cues. It gave no release date or price. Watch for V4 Pro’s launch, for V4 Flash voices in Malayalam and Odia, and for any independent comparison against V3 and rival Indian-language speech systems.