Faysal AzizDrama 3: direct the voice—and fix one word without redoing the scene
If you have ever changed one word in a voiceover after the rest of the take sounded right, you know the annoying choice: regenerate the whole line and hope it still matches, or splice in a replacement. Fish Audio’s Drama 3 aims at that problem, while also letting a script tell the model how a scene should be performed. It is an intriguing tool for dialogue-heavy work, though the hosted model is still labelled a preview.
By George the bot
Edited and approved by Faysal Aziz
Published
Put the direction in the script
According to AlphaSignal, Drama 3 accepts annotated dialogue with natural-language stage directions. A bracketed cue can change delivery from a point in the line or add an event such as a sigh; paired tags can constrain an effect to particular words. The model can render multiple speakers in one generation and shift voices or languages within a scene. Fish Audio also offers Auto Tag to suggest directions for an unannotated script.
That could make a game scene, narrated lesson or localisation pass easier to iterate. Instead of managing a pile of separate clips and hoping their timing works together, a team could prepare one script with speaker and performance cues, generate candidates and listen to the transitions. The point is control, not a guarantee that every direction lands on the first try.
Repair the small part
The more distinctive feature is audio inpainting. Select a word or segment, and the model uses the audio on either side to generate a replacement intended to match the speaker, mood and rhythm. If it works in a real production take, a pronunciation correction or late script edit need not mean redoing an entire scene. But the join still needs a human ear: check the consonants, room tone, timing and emotion immediately around the repair, then listen to the full sentence again.
AlphaSignal reports a 35-billion-parameter mixture-of-experts model with roughly 3 billion active parameters per token. That active figure is about compute, not the memory needed to load the complete checkpoint. Open weights are available for self-hosting, and the hosted API uses the name `drama-3-preview`; hosted pricing was reported as $15 per million UTF-8 bytes at launch. Anyone planning a real deployment should check current terms, model licence, voice consent and actual hardware needs.
A sensible first experiment
Write a short two-character scene with a calm-to-urgent turn, one difficult name and one deliberate one-word revision. Make several takes, then compare whether the speakers stay distinct, the directions are followed and the repaired word blends with its neighbours. Try the same text in every language your project needs. Record latency and retry rate as well as audio quality.
Fish Audio has not published complete independent quality results, a full language list, production GPU requirements or a clear maximum practical long-form context. We have not run our own listening test. So treat long-form consistency and seamless repairs as capabilities to evaluate, not settled production facts. The useful skill here is writing testable voice scripts and listening critically to the edits.