Custom Pronunciations lets you set how specific words should be said, on a per-project basis. It is built for out-of-vocabulary terms: brand names, acronyms, and domain-specific words the model might otherwise mispronounce.

Each override you define syncs to the cache automatically and is resolved in under 10ms at synthesis time, so adding pronunciations does not slow generation down.

Because the overrides are scoped per project, different projects can pronounce the same word differently, and the terms that matter to your brand get read correctly every time.