Sarvam ships Bulbul V4 Flash for high-volume Indian voice agents

Indian AI lab Sarvam has released Bulbul V4 Flash, a text-to-speech model designed for Indian languages, regional accents, Hinglish code-switching, identifiers, acronyms, and currency amounts. The company distilled Flash from a larger V4 research model to reduce inference cost and latency for voice agents handling large numbers of concurrent calls.

Sarvam made Flash available through its API playground, Content Studio, and the existing Bulbul API endpoint. Current Bulbul customers can adopt it by changing their model selection instead of rebuilding the integration.

The launch in numbers

The launch announcement also advertised a temporary 50% promotion that reduced the price to ₹10 per 10,000 characters through November 12. Developers integrating after the promotion should confirm the current rate in Sarvam’s console.

Distillation shrinks the serving bill

Sarvam’s V4 research program removed production constraints on model size, latency, and cost while the team tested larger language backbones, richer audio representations, multiple generative approaches, and more than 400 billion audio tokens. Audio tokens are compact machine-readable units used during training, so that figure cannot be converted directly into recorded hours without details about the tokenizer and dataset.

Distillation transfers useful behavior from a large teacher model into a smaller student model that is cheaper to serve. Sarvam says its experiments identified which components improved pronunciation, prosody, contextual delivery, and overall speech quality, then compressed those gains into Flash.

At list price, Flash costs one-third less than V3 while Sarvam reports better speech quality and lower latency. Those economics matter in call-center deployments, where small differences in generation cost accumulate across millions of characters and concurrent streams.

Indian speech stays central

V4 was pretrained from scratch with Indian-language representation as a primary objective. Sarvam used a continuous, language-balanced training curriculum intended to preserve native pronunciation, speaker identity, and expressive delivery across the training run.

According to the company, its proprietary speech corpus grew nearly sevenfold from V3. The expanded data covers additional languages, regional accents, speakers, emotional registers, and conversational styles.

A final specialization stage tunes each voice for its accent and conversational rhythm. That process aims to prevent a compact multilingual model from averaging away pronunciation patterns that distinguish Indian languages and Indian English.

Normalization handles awkward tokens

Flash uses a real-time contextual normalization layer to determine how text should be spoken before synthesis begins. The layer classifies acronyms, identifiers, phone numbers, dates, and currency amounts according to their role in the surrounding sentence.

• ISRO is pronounced as a word.
• EMI is read letter by letter.
• PAN is pronounced as a word when it names the document, while the associated identifier requires character-level handling.
• Aadhaar receives its familiar local pronunciation.

Expressive alignment during training connects those normalized forms to the generated delivery. In a banking or travel call, accurate verbalization of a payment amount, PNR, phone number, or account identifier can reduce repetition and correction turns.

Concurrency drives the serving design

Time to first audio measures the interval between sending text and receiving the first playable audio chunk. End-to-end voice-agent latency also includes speech recognition, language-model generation, network transit, buffering, and playback, so the reported sub-250 ms figure covers only the synthesis stage.

• Receptive-field-aware streaming skips computation that cannot affect the next audio segment.
• Real-time pacing prioritizes audio needed by active conversations instead of generating far ahead of playback.
• Reduced synchronization and copying lowers overhead between inference operations.
• Grouped streaming work batches compatible requests without blocking active streams.
• Live-load balancing distributes requests according to current server utilization.

Under comparable load, Sarvam says these changes make Flash up to 40% faster than V3 and leave more capacity for concurrent conversations. The published figure remains vendor-reported, and teams should test it against their own request lengths, concurrency patterns, network regions, and latency percentiles.

Voices, styles, and boundaries

The launch catalog includes 83 voices across 11 languages and 224 voice-and-style combinations. Sarvam targets customer care, sales, collections, banking, e-commerce, education, narration, and general conversational applications.

Style controls allow one voice to support several roles, such as a collections agent, tutor, or sales representative. Developers should confirm that their required language, voice, and style combination appears in the API catalog before designing a fixed persona around it.

Sarvam is also developing Bulbul V4 Pro for applications requiring wider paralinguistic control, including laughter, pauses, sighs, excitement, nervousness, and precise emphasis. Flash prioritizes cost, latency, pronunciation, and conversational delivery, while Pro is intended for storytelling and emotion-heavy experiences.

The announced scope centers on Indian languages and Indian English. Products that also require European or East Asian voices may need another model or provider, adding routing, monitoring, and voice-consistency work.

Workloads that fit

• Banking, lending, collections, and insurance agents that must pronounce numbers, names, and legal terms accurately.
• Railway, airline, and logistics helplines that combine Hinglish with booking references and identifiers.
• Education and customer-support bots that require Indian English or regional-language voices.
• Interactive voice response systems whose unit economics depend on low per-character synthesis costs.
• High-concurrency call operations that need streaming audio with predictable time to first byte.

Ship with a regression suite

Existing Bulbul clients can retain the current API surface and change the selected model, but production migration still requires output and load testing. A practical rollout should cover the following checks:

1. Build a pronunciation set containing customer names, addresses, currency amounts, dates, acronyms, phone numbers, PANs, PNRs, and mixed-language sentences.
2. Compare voice and style identifiers with the V3 configuration and verify that the intended persona remains available.
3. Measure median and tail latency at realistic concurrency, including p95 and p99 time to first audio.
4. Track stream interruptions, audio gaps, retries, and cancellation behavior during long calls.
5. Measure the complete conversational turn, including speech recognition, model generation, synthesis, network transit, and playback buffering.
6. Confirm current pricing, character-counting rules, quotas, and regional availability before estimating per-call cost.

All performance and quality figures in the launch material come from Sarvam. Teams evaluating Flash should pair the straightforward V3 migration with their own listening tests, pronunciation benchmarks, and production-scale load tests.