Fish Audio | Drama 3 Adds Expressive Direction to AI Voiceovers
Fish Audio released Drama 3 on October 13, 2026, focusing on expressive text-to-speech for scripted performances. It is available through the website, iOS and Android apps, and API. For a video editor, the appeal is directing a change in delivery where the scene needs it, rather than treating the whole script as one emotional setting.
Direct the moment when the performance changes
Drama 3 accepts natural-language direction in square brackets, including instructions placed partway through a line. It supports 83 languages, and the language of the direction can differ from the spoken text. This offers a practical way to describe an emotional turn without rewriting the dialogue itself.
For an illustrated short, identify the line where the character's intention changes. An apparently confident introduction could give way to hesitation, or a restrained answer might end in relief. Decide what the audience should understand at that point, then compare whether the generated delivery communicates it alongside the visuals.
Evaluate emotion and intelligibility together
Prepare a short test that includes both an emotional beat and words that must be unmistakable. A name, a date or a product term can reveal a different problem from a flat performance. Listen without reading the script first, then compare against the wording to catch omissions or unclear pronunciation.
Review the take within the actual edit. A pause that sounds effective alone might arrive too late for a cut; a laugh could obscure the final phrase beneath music. Mark the moments that need revision and preserve the parts already serving the scene. The goal is a coherent performance, not maximum emotion in every sentence.
Separate hosted access from open-weight deployment
Drama 3 is also distributed as open weights. The Fish Audio Community License allows qualifying commercial use below a US$500,000 annual-revenue threshold, including affiliates; larger organizations and certain hosted-service uses require separate permission. Voice cloning without the person's consent is prohibited. These conditions should be checked before choosing a deployment route.
A creator evaluating the hosted tool can begin with an existing production task rather than a deployment project. Compare voice suitability, timing and the effort needed to reach an approved take. If local operation is important, review its setup and license separately from the decision about which performance belongs in the video.
Build a listening reference for a series
For recurring narration, keep a few approved passages that establish the expected pace and personality. Include a calm explanation, a sentence with specialist wording and a closing line. Use those examples to judge later replacements, especially when only a small part of a finished sequence needs changing.
For multilingual material, have a fluent reviewer assess the meaning and delivery in context. Compare each version with its own on-screen timing rather than forcing every language into the same rhythm. A consistent identity should still leave enough room for the speech to feel natural and remain easy to follow.
Daisuki's Take: What This Means for Designers
We see performance direction as an editorial skill worth carrying into generated speech. The best instruction describes the scene's intention rather than asking for a generally dramatic voice.
We would test a difficult line first: one that combines precise wording with a change in mood. That makes the model's usefulness easier to assess before a long narration.
For creators, a successful take supports the story and survives a careful listen. Expressiveness matters most when it strengthens the message without losing the words.
Sources and Recommended Links
- Fish Audio Drama 3: The Most Emotive Text-to-Speech Model We've Built | Fish Audio Official Blog
- Fish Audio Drama 3 model card | Fish Audio Official Model Repository
