The thesis is: Natural-language and word-scoped directions make performance control an explicit part of a TTS script, rather than a separate post-generation adjustment. Fish Audio’s product page shows directions for emotion, intent, pacing, and addressee, alongside inline pauses and sounds. It also shows effects scoped to selected words, such as whispering or slower delivery.
The mechanism is: Script-local directions and word-scoped tags provide controls within the text being synthesized; the API accepts the same direction format. The page also describes multi-speaker scripts generated together and API requests using the drama-3 model header. Why now: Drama 3 is presented as available now in Fish Audio’s Text to Speech product and callable through its API, so the directing interface can be tried in both the editor and an integration.
The first implication is: For teams building voiced dialogue, direction can be authored and revised alongside the words instead of treated only as a global voice setting. That makes script structure a practical control surface, although it does not establish consistency or production quality. The evidence boundary is: The supplied product page demonstrates the interface and access path, but gives no independent quality benchmark, cost comparison, or proof of open weights. Fish Audio says Drama 3 remains in Preview and results vary between takes. The first watch condition is: Watch whether later documentation adds reproducible quality or consistency evidence and clarifies the model’s status beyond Preview.