Fish Speech Drama 3: A Dialogue-Editing Preview

Illustrative fictional voice performer recording dialogue; not a Fish Drama 3 test output.

Fish Speech Drama 3 raises a useful question for dialogue creators: can you repair one phrase without discarding a performance? Keep the access distinction in view. Fish announced ordinary Drama 3 text-to-speech for its web, app and API, while interactive editing and audio inpainting were described as small invite-only tests. A research capability is not a button every customer can use today.

For an episodic fictional story, follow the micro drama creation workflow to connect character references, planned shots, dialogue and the final cut.

Prepare an approved dialogue take for a separate APOB lip-sync pass.

This preview explains the research and supplies an original repair specification. We have not run an invited editing session, produced repaired audio or measured preservation around a splice. If your account lacks editing access, the practical fallback is an authorized full-line retake and an external edit. Keep this specification for evaluation rather than pretending an unavailable interface exists.

Locate Drama 3 within the Fish Speech project

The October 13 official release

The Drama 3 technical report is dated October 13, 2026. It describes speech generation research including revision across conversational turns and infilling within existing speech. Those are the relevant concepts here; this article does not reproduce a benchmark table or infer superiority over another service.

Check the current Fish announcement and your account before making a production commitment. Availability can change after a release. An announcement about TTS, an invited editing experiment and an editing API described as forthcoming must remain separate facts in your brief.

Open model research versus hosted interfaces

The report links the Fish Speech project and model resources. That research distribution is a different surface from a hosted customer editor. A model being available for research does not tell you what controls your current hosted account exposes.

Use the official Fish Speech repository for project information. Do not assume its installation instructions, model permissions or inference requirements are the same as a commercial account's workflow. No licensing clearance is implied by the writing exercise below.

What this article has and has not tested

Our materials are a marked transcript, three proposed direction notes, an empty listening log and a downstream retake brief. They are authored planning assets. There is no audio sample, panel score, join-quality result or successful local repair behind them.

That boundary matters when deciding what to keep. “Protected context” below describes the reviewer’s requirement. It does not claim that any model preserves neighboring samples exactly. You must compare actual returned audio before accepting it.

Translate the research terms without inventing an editor interface

Revision after spoken feedback

Revision asks for another attempt in response to direction. In our example, an adult stage manager says a rehearsal time correctly but stresses it sarcastically. A note might ask for a sincere delivery without changing the words. The requested change concerns performance.

The report's discussion of multi-turn revision supports investigating that operation. It does not justify inventing a “remove sarcasm” toggle or documenting an endpoint that the announcement does not make public.

Infilling inside an existing utterance

Infilling addresses a bounded part of an existing utterance. Our proposed example changes “seven” to “eight” in a rehearsal instruction. The surrounding sentences are material the reviewer wants retained and compared, not disposable padding.

Proposed operation

Original example

Question for a future authorized trial

Revision

Same words, less sarcastic emphasis

Does the new delivery fit the intended relationship?

Infilling

Rehearsal time changes from seven to eight

Are words and transitions acceptable around the change?

What a research result does not establish about account access

Neither row is a UI procedure. Before a trial, record the actual service, model, access level and operation available to your account. If the editor is not accessible, stop describing the repair as something already performed.

A normal rerecording can still solve the production problem. The relevant comparison is between a returned, inspected local repair and a returned, inspected full-line take. An unavailable experiment cannot win that comparison on the strength of a research description.

Fence the editable phrase with protected context

The last unchanged word before the edit

Authored transcript, two fictional adults:

Stage manager: “I checked the room. Rehearsal starts at [CHANGE: seven → eight]. Please use the side entrance.”

Performer: “Eight, then. I'll tell the others.”

The final unchanged word before the target is “at.” Keep the original audio and transcript together so a reviewer can listen to the transition rather than inspect the replacement word in isolation.

The first unchanged word after it

“Please” begins the next sentence. It matters because a time correction can change the pause leading into that instruction. A transcript that contains the right number may still sound as though two unrelated recordings were joined.

These brackets are editorial notation, not API syntax. Do not paste them into an interface while assuming they activate protection. Translate the requirement into whatever controls an authorized trial actually provides, or use them to brief a human rerecording.

Speaker, room tone and emotional momentum

Record the intended speaker and permission basis, the room-tone character and the scene's emotional direction. Here the manager is correcting a schedule, not escalating a conflict. A louder or accusatory “eight” would alter the exchange even if it were intelligible.

For performance preparation, the existing voice-acting guide addresses a different job. This article concentrates on specifying and reviewing a proposed change to an already chosen take.

Write a note that changes only the intended dimension

Correct the time without making the line louder

Director note A: “Replace the rehearsal time with eight. Keep the surrounding instruction conversational. Reject the attempt if the corrected word becomes an emphatic reprimand.” Allowed change: factual time. Protected requirement: the manager's calm intention.

Do not add “make it more emotional” to the same note. That creates a second variable and makes it harder to identify what changed. If both fact and performance need repair, state them separately and review both against the original.

Keep the apology but remove the sarcastic emphasis

Director note B: “Keep ‘I'm sorry about the change’ and its meaning. Ask for a sincere, restrained reading. Reject a version that drops the apology or turns it into a joke.” This is a proposed revision note, not an assertion that ordinary accounts can edit an existing Fish take today.

The report describes research relevant to repeated direction. Keep that claim distinct from the hosted availability statement. The technical report is the source for the former, not proof of the latter.

Reject a repair that adds an unscripted breath

Director note C: “Keep the transition into the next sentence unobtrusive. Reject a new breath that makes the manager sound hesitant before the entrance instruction.” The rejection condition makes the note reviewable.

An added breath is not universally wrong. It is wrong for this specified intention if it changes what the listener infers. Preserve the unsuccessful attempt as evidence of that mismatch instead of quietly revising the brief after hearing it.

Listen to the join before trusting the repaired words

A clipped consonant at the entrance

Listen to the preceding word, the target phrase and the following sentence in context. A hypothetical failure would clip the end of “at” while delivering a clear “eight.” This is an invented example of what to check, not a defect observed in Drama 3 output.

Use headphones and ordinary playback conditions appropriate to the intended release. Note what you actually hear; avoid claiming sample-perfect preservation from a casual listen or a matching transcript.

An abrupt room-tone change

Listening-log field

Fill only after an actual trial

Original take and permission record

___

Service, model and access surface

___

Requested phrase or performance change

___

Entry/exit transition observation

___

Room tone and speaker continuity

___

Accepted, rejected or uncertain; reason

___

The blank cells are deliberate. A checklist is not a result. If no authorized output exists, keep them empty and describe the item as pending.

The next speaker responding to the wrong emotion

Play the performer’s reply after the candidate repair. A new sarcastic manager take could make a calm reply feel implausible. That is a scene-level mismatch, even when both recordings sound clean on their own.

Accept the audio only when the factual change, transitions and relationship all fit the approved brief. If a full-line retake provides a more usable result, use it. Local editing is a possible production tool, not an obligation to preserve an unsuitable original performance.

Reopen lip-sync only where the accepted audio changed

A new mouth-performance pass for the affected shot

Replacement brief: affected shot ___; original audio ___; accepted replacement ___; approved fictional character reference ___; required line ___; unresolved visual check ___. Preserve the original and identify the accepted take explicitly.

APOB lip sync is an optional next production step for supported authorized audio and visual inputs. Its practical role is generating a new speaking performance to inspect. It is not an automatic Fish repair connector and does not guarantee identical neighboring frames.

An unchanged reaction shot that may remain usable

An unrelated listening shot may survive the audio change if its action, eyeline and duration still fit. Check it rather than replacing every asset by default. A retained shot is a production decision, not proof of a cross-tool dependency system.

If you need a new visual source, investigate the APOB image-to-video route against its actual model support. Keep asset creation separate from claims about speech-editing access.

A separate final-edit timing check

An external editor must review the accepted audio, speaking shot, reaction and subtitle timing together. The new phrase may alter the usable duration. Matching words alone cannot certify the assembled scene.

As of November 2, 2026, this remains an access-aware research preview and an original repair specification. Verify Fish's current availability before a trial. Where invite access is absent, use an authorized full-line retake; do not invent the missing interface or describe these unfilled checks as completed tests.

Sources

HappyHorse 1.1 vs HappyHorse 1.0: Feature Analysis for AI Video GenerationAPOB AI generated AI imageSoft gradient background of the APOB AI call-to-action banner

Create a dreamlike vision with APOB

No credit card needed

CREATE FOR FREE

You may also be interested in