Explainer scripts get rewritten after the first cut. Generating line by line means a late change costs one line rather than a re-record, and audio arrives in about a second at 44.1 kHz.

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.

Podcast production is segment based already, which suits generation in short calls. Intros and ad reads generate once and cache, and pairing contrasting voices produces a two-hander without booking two people.

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.

One hundred and forty six voices are tagged young, 63 percent of the catalog, across every accent and all twelve recommended languages. Unlike tone, age is a real tagged field rather than an inference.

Only six voices in the catalog show cheerful or positive markers, the thinnest tone signal available. Speed slightly above 1.0 lifts almost any voice, which is the more dependable route.

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.

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.

Announcements are heard in reverberant halls and repeated thousands of times. Pronunciation dictionaries fix station and place names permanently, ulaw and alaw feed public address hardware directly, and one voice carries a full multilingual chain.

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.