Voice agents live or die on the pause before a reply. Synthesis starts in about 200 milliseconds and runs at 3.3 times real time, which leaves the budget where it usually belongs: your model.

Support calls reach people who are already inconvenienced, often on a poor line. These voices favour clarity over character, hold steady across renders, and move between Hindi and English the way callers actually do.

IVR voices need to survive a compressed phone line, not just sound good in a browser. These output ulaw and alaw directly, generate in roughly 200 milliseconds, and handle Hindi and English in the same prompt.

One hundred and ninety five voices carry the conversational tag, 84 percent of the catalog. Streaming is what makes them feel live: audio begins playing before the sentence has finished rendering.

For Indian customer facing work the deciding factor is not tone but language. These voices switch between Hindi and English mid sentence, which reads as considerably more natural than either alone.

Seventy two voices cover eleven Indic languages plus English, so one voice serves a Hindi caller and a Tamil caller without re-casting. Hindi and English can alternate inside a single sentence.

Generation runs at 3.3 times real time, so audio renders faster than it plays and the buffer stays ahead. First audio arrives in about 200 milliseconds over SSE or WebSocket.

At volume the cost driver is duplication, not generation. Hashing text and skipping what already exists removes most of it, since catalog and template work repeats the same phrases constantly.

Pre-rendering every branch means shipping every branch. WebSocket streaming lets dialogue generate during play instead, so a conversation tree costs what players actually hear rather than what they might.

Guided audio lives on pacing rather than voice character. Speed goes down to 0.5, pauses come from how you write the script, and nine Indic languages are available for regional wellness content.