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.

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.

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.

A novel is thousands of requests stitched together, and the joins are where narration falls apart. These voices hold consistent across a full book and cover nine Indic languages that most vendors do not.

Speed is the reliable lever here, not casting. Settings between 1.2 and 1.4 lift almost any voice in the catalog, though longer sentences start to blur past roughly 1.5.

Slowing down reads as more authoritative than speeding up. A setting near 0.9 does more than any voice choice, and short declarative sentences carry more weight than qualified ones.

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.

Web chat is unusual because the visitor reads and listens at once. These voices suit an unhurried delivery, begin playing in about 200 milliseconds over WebSocket, and cover twelve languages from one integration.

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.

Commercial reads are short, fast, and worth testing in variants. A thirty second spot is one or two requests, which makes producing four versions cheaper than booking one session.