Fish Audio launches Drama 3, a voice AI capable of shifting emotions mid-sentence

Fish Audio launches Drama 3, its new open-weight voice model capable of controlling emotions, intonations, and the pronunciation of every word across 83 languages.

A voice cracking in the middle of a sentence, an uncontrollable laugh, anger expressed through clenched teeth. With Drama 3, Fish Audio aims to give creators the ability to direct a vocal performance with the precision of a director. Its new text-to-speech model does not merely read text anymore: it allows users to fine-tune emotions, breaths, and intonation shifts down to a single word. Directing a voice like an actor

Drama 3 relies on natural language instructions placed directly within the script in square brackets. Simply specifying that a character is holding back tears, whispering, or bursting into laughter is enough to guide their performance. Directives can change mid-line without altering the character's voice.

The model also supports tags that allow users to emphasize certain words or modify their pronunciation. This approach is particularly compelling for audiobooks, audio fiction, animation, and video games, where a single line can shift from one emotion to another in a matter of seconds.

Drama 3 works in 83 languages, and the instructions do not need to be written in the dialogue's language. A French-speaking creator can thus direct an English line using prompts written in French. Better emotional control without sacrificing words

This expressiveness goes hand in hand with work on reading accuracy. Drama 3 achieves 93.2% correct pronunciations, compared to 73.2% for S2.1 Pro, on Artificial Analysis's pronunciation robustness benchmark. This independent test covers challenges specific to English, including numbers, abbreviations, names, and technical sequences.

Fish Audio has also modified its training method. Rather than prioritizing only fidelity to a reference voice or ease of transcription, the model is fine-tuned based on human preferences regarding the generated performances. A laugh requested in the script is no longer penalized simply because it disrupts the conventional reading of the text. A 35-billion-parameter open-weights model

Drama 3 uses a Mixture-of-Experts (MoE) architecture of 35 billion parameters, with approximately 3 billion active during each generation step.

Fish Audio is also releasing the model weights, allowing execution on one's own infrastructure. However, this openness comes with a specific license: commercial use is free for entities whose annual revenue, including affiliates, remains under $500,000. Beyond that threshold, or to operate the model as a third-party service, an additional license is required. Modifying a word without re-doing the entire line

Fish Audio is preparing two additional features, currently accessible only within limited testing. The first will allow users to request a new performance based on an existing take, for example by making a sentence softer or more ironic.

The second, based on targeted audio editing (inpainting), will allow a single word in a recording to be replaced while preserving the rest of the line, its rhythm, and its intent.

Drama 3 is already available on Fish Audio, in mobile apps, and via the API. An 80% discount on the API is offered until November 12, as well as a 50% discount on new creator subscriptions.

With this model, Fish Audio addresses a dimension that remains delicate for speech synthesis: voice direction. For audio professionals, the challenge is no longer limited to producing a believable voice, but extends to being able to adjust an intent, correct a performance, and keep successful takes, just like in actual studio work.