Fish Audio: Drama 3

Fish Audio: Drama 3 is a text to speech model from Fish Audio, listed on api14 under the speech category. It is built for directed voice performances, which means the model is designed around the idea that you describe how a line should be delivered rather than just supplying the words. Fish Audio is the provider behind the model, and the provider release date is 13 October 2026.

The model takes text as its only input and returns speech as its only output. Its distinguishing capability is direction handling: you write the direction in plain language inside square brackets before a line, for example a bracketed note such as nervous before the sentence that should be spoken nervously. That lets you shape the performance from the prompt itself instead of relying on separate controls. No context length is published for this model, and the details list a maximum output of zero, so there is no documented token window or output token budget to plan against. It is a speech generation model, not a text or multimodal model, so it does not accept audio or image inputs and does not return text.

The practical uses follow from that design. Because direction is written in plain language, Drama 3 suits work where delivery matters as much as the words: character dialogue and voice acting for games, animation or audio drama; narration for stories and podcasts where tone shifts between lines; and advertising or explainer voiceovers where a single script needs several emotional readings. It also fits prototyping, since you can iterate on a performance by editing the bracketed direction rather than re recording a human voice actor. Any workflow that starts from a script and needs spoken output with per line control is a reasonable candidate.

You call the model through api14's OpenAI compatible API using the model id "fish-audio/drama-3". Because the interface follows the OpenAI format, you send a request to the speech endpoint with that id and your text input, and you receive generated speech in the response. The model lists no supported parameters, so there are no documented sampling or generation options to set beyond the model id and the input itself; treat the request as the model id plus your directed text. If you need to confirm exact request and response fields, check the endpoint schema on the model page.

There are limits worth knowing before you commit. No context length is published, and the maximum output is listed as zero, so there is no stated ceiling on how much text you can send or how much speech you can get back in one call. Supported parameters are empty, so if your pipeline depends on tuning generation settings, this model does not expose them. If your task is text generation, transcription or anything other than producing speech from text, a different model in the catalogue will fit better.