Thesis: In real-time speech products, model compression becomes valuable when it preserves language-specific delivery while lowering the cost and delay of serving. The announcement reports under 250 ms time-to-first-audio in production and a price of ₹20 per 10,000 characters, compared with ₹30 for Bulbul V3.
Mechanism: Sarvam says it distilled gains from larger V4 research into Flash, optimized the inference stack, expanded its proprietary speech datasets nearly sevenfold relative to V3, and trained with a continuous language-balanced curriculum followed by voice specialization. It also describes contextual text normalization for names, numbers, dates, and reference codes. Why now: The company frames fast, predictable audio response as necessary for real-time conversations, where pauses can disrupt interaction and per-character costs accumulate with usage.
First implication: Teams evaluating multilingual voice agents can compare latency and cost alongside pronunciation, prosody, and handling of names and numbers in their actual target languages. Evidence boundary: These figures and descriptions come from Sarvam's announcement; it supplies no independent benchmark or comparative evaluation in the cited material, so they should be treated as vendor claims rather than verified third-party results. First watch condition: Independent tests should establish whether sub-250 ms time-to-first-audio and the stated price hold under representative production loads and language-specific quality checks.