Indian-language speech synthesis drops under 250 ms: 83 voices across 11 languages, at a third less cost
Synopsis
Sarvam released Bulbul V4 Flash, a text-to-speech model for Indian languages that reaches under 250 ms time-to-first-audio in production and costs ₹20 per 10,000 characters. That is a third less than V3's ₹30, while the company reports richer prosody, stronger contextual delivery and higher speech quality. The model launches with 83 voices across 11 languages and 224 voice-and-style combinations, plus a real-time contextual text normalization layer for items such as PNR, PAN and Aadhaar whose spoken form depends on context.
Interpretation
Real-time speech synthesis for Indian languages can now reach under 250 ms time-to-first-audio in production while keeping expressiveness. The previous Bulbul V3 was priced at ₹30 per 10,000 characters; V4 Flash costs ₹20, a third less, and the company reports richer prosody, stronger contextual delivery and higher speech quality. Under comparable load, V4 Flash is up to 40% faster than V3 with substantially more headroom for concurrent conversations, measured in production.
The model compresses multilingual speech expressiveness into a smaller model while keeping the qualities listeners notice most. The V4 research phase set aside production latency and cost limits, trained larger language backbones, explored richer audio representations and scaled training data beyond 400 billion audio tokens. At that scale, more data and more capable models kept improving the model's use of context, prosody, expressiveness and robustness, and those gains were then distilled into the smaller Flash model.
The inference stack organizes computation as a stream, skipping work that cannot affect the next stretch of audio. The stack uses receptive-field-aware streaming to skip work that cannot affect the next stretch of audio, and real-time pacing keeps compute on what active conversations need now. The serving path also cuts unnecessary synchronization and copying, groups compatible streaming work and balances requests by live load, with every optimization gated on output correctness.
Perspective
The report covers the Bulbul V4 Flash release as announced by Sarvam: under 250 ms time-to-first-audio in production, ₹20 per 10,000 characters, 11 languages, 83 voices, 224 voice-and-style combinations, and up to 40% faster than V3 under comparable load. These figures target developers and enterprises that need real-time speech, available through the Sarvam API, Playground and Content Studio, with existing Bulbul users able to migrate on the familiar API.
The company reports that quality had not plateaued at more than 400 billion audio tokens, but Flash is a distilled smaller model, and what it keeps is bounded by the qualities listeners notice most. The latency and speedup figures come from production and comparable-load conditions, with no independent reproduction or public benchmark. V4 Pro appears only as a preview and has not been released.