giving every character their own voice

Voice binding solves one thing: making a character sound the same every time they open their mouth.

Bind a recording to an actor and, from then on, every shoot segment that brings that character in carries the voice automatically — no picking it per segment, no writing it into the prompt.

This handbook covers two things: how to bind one, and where it stops working. The second matters more — voices come out of a capped budget, and you want to know the limits before a finished shot tells you someone's voice never made it.

The fast path: three steps

1. Find the actor — left sidebar, "Create → Casting". The badge row under each actor card has a "＋ Voice" badge that takes you straight there; or open the casting room and press "Edit identity", where "Voice" is one of the sections.

2. Give it a recording — two sources:

• "Upload a voice" — your own audio file.
• "Pick from the studio" — anything you generated in the Recording Studio, plus audio you uploaded to the Library.

3. Done — the row then shows the file name, its length and a play button. The badge on the actor card changes from "＋ Voice" to "Voice".

Two other entry points write to the same place: character cards in the Library / art station have a "Bind voice" button, and every clip in the Recording Studio has "Bind as a character voice". All three bind the same thing.

The rules

It binds to the person, not to the picture

The voice hangs off the actor, not off one image. So:

• Swapping the main image, editing it again, pressing "↻ Regenerate" for a new face — none of that breaks the binding.
• Every look for that actor (makeup, wardrobe, covers) shares the one voice.
• Copies made when you add the actor to the Library or promote them into a project are recognised too.

The flip side: one actor, one voice. If the same face needs a different voice in a different episode, make it two actors.

What the file has to be

Files that do not qualify are rejected at upload — far easier to understand than a shoot that fails later.

Why only five seconds

Give it 30 seconds and the model still receives 5. That is not us cutting corners:

• The recommended range is 5–10 seconds per clip; shorter or longer both reduce accuracy.
• More importantly, a scene has a total audio budget. At a fixed 5 seconds per actor, "how many voices fit in this scene" becomes a number you can count (see Limits below). One 30-second clip would eat a whole scene's budget by itself.

Leading silence is trimmed first, then five seconds are taken from where the speech actually starts.

⚠️ After the leading silence is trimmed there must still be a full five seconds of speech, otherwise the binding is refused:

A character voice needs at least 5 seconds of actual speech once leading silence is trimmed — use a longer recording.

So "the file is 8 seconds, but the first 4 are empty" gets rejected. A file passing on length does not mean the sample passes.

What happens automatically at shoot time

1. Whoever the segment brings in gets their voice. That includes voice-over characters (they speak but their face is not on screen), and they are queued ahead of on-screen characters.
2. The anchor lines are appended to the prompt for you, telling the model which clip belongs to whom. You never write @Audio 1 yourself.
3. A character with a bound file gets the file, and their text voice description is not sent. Characters without one use the text description. It is one or the other, not both — see limit ⑥.

Limits

① How many voices fit in one scene

Audio has a total-length budget and each actor costs five seconds, so the headcount is fixed:

Anyone over the line does not vanish — they fall back to their text voice description (the field in the acting profile; see the Casting Handbook). So for a crowded scene: bind files for the leads, and write everyone else's text description.

Queue order (earlier gets the slot):

1. Reference audio you attached to that segment by hand
2. Voice-over characters (they definitely speak)
3. On-screen characters

⚠️ Manual tracks eat the budget first. Attach a 15-second music bed to a Seedance 2.0 segment and there is nothing left for voices.

② You can see who makes it before you shoot

There is a chip above the generate box on the shoot page:

Voices 4/6

Open it and it lists who was sent and who fell back to a text description. Switching engine, adding or removing characters, attaching another audio track — the number recalculates immediately.

Nothing drops silently — you see it before you spend anything. The chip only appears for people who have bindings.

⚠️ One easy-to-miss knock-on: attaching one more reference image by hand pushes an automatically selected character out of the segment, and their voice goes with them. The chip's number moves, which is how you spot it.

③ First/last frame carries no voices at all

First/last frame (reference image 1 becomes the opening frame, image 2 the closing one) and reference audio cannot coexist — that is a model-side restriction, not our choice.

So that segment falls back to text descriptions for everyone. Voices or first/last frame: pick one.

④ The Studio (quick shots) does not use bindings

This one gets misread the most.

Pick a cast actor's image in the Studio and, as far as the system is concerned, it is just a reference image — it cannot tell which actor that is, so voice bindings never enter the picture. The Studio only sends the one audio track you attached by hand.

Voice bindings apply to shoot page segments (short films, short dramas), because those segments carry real character assets.

⑤ It is a reference for vocal quality, not dubbing

This sets your expectations for the result:

• It can — have the model take this person's timbre and delivery and speak the lines in the scene.
• It cannot — play your recording as narration, or have the model speak new lines that are not in the recording in that voice. Quality drops noticeably there.

Bound does not mean a 100% match. The vendor guidance says it plainly: you can upload audio for the model to imitate, but audio alone often falls short of what you expect. It is a strong steering wheel, not a copier.

⑥ File and text are one or the other, never both

This is the easiest limit to misread, and the cost of misreading it is concrete.

For any one character in any one segment, only a single track is used:

Which means: the moment you bind a file, you have chosen the sample over the description. They do not add up; the text stands down. (The rest of the acting profile — habitual and signature behaviour — is unaffected and always goes.)

So which one should you use?

Whichever you choose, write the text description anyway — it is never wasted. It covers the segments with no slot left, the segments with first/last frame on, and the day you unbind the file.

Replacing and removing

• Replace — "Replace" at the right of the row binds a new one; you do not need to remove the old one first.
• Remove — the ✕ next to it. Removing does not need a paid plan and can be done any time. Once removed, that actor's text voice description takes over again (limit ⑥).

Changing the voice does not change shots you have already generated — bindings only affect what is sent from then on.

Where things are

Credits and plans

• Binding and unbinding cost nothing. The voice only rides along with a video generation; it is not billed separately.
• Binding is a paid-plan feature. Free plans do not see the Voice section and cannot add or replace a binding — but an existing one can still be removed.
• Generating audio in the Recording Studio does cost credits; the button says how many.

Common questions

I bound a voice and the result still does not sound like it.
Check three things first: ① did the shoot page chip count them under "Sent"; ② does that segment have first/last frame on (that carries none at all); ③ was it generated in the Studio (the Studio ignores bindings). If all three are fine, it is a question of imitation strength — a file is never a guaranteed match (limit ⑤). At that point unbinding and switching to a text description usually beats trying yet another recording, because the two are never sent together (limit ⑥).

Would writing both make it more accurate?
No — for one character in one segment only one of them is used, and a bound file wins. Write the text description anyway: it is the fallback when there is no slot, when first/last frame is on, and when you later unbind.

Five characters in the scene and only the first few have voices.
That is the budget, not a fault. Three per scene on Seedance 2.0, six on 2.5. Switch to 2.5, drop a manual audio track, or move the rest to text descriptions.

My upload was rejected for length.
It has to be 5–30 seconds, wav or mp3. If it is inside that range and still rejected, the leading silence is usually too long — the clip is measured from where the speech starts, and a full five seconds has to be left.

I changed the face / the main image. Do I rebind?
No. The binding follows the actor, not the image.

I want two voices for one actor.
One voice per actor today. Split them into two actors ("young" and "older", say) and bind one each.

Do already-generated segments change?
No. Bindings only affect what is sent afterwards; reshoot the segment to update it.

I cannot find the voice I want in the studio picker.
That list is your own Recording Studio output plus audio you uploaded to the Library. Generate one in the studio, or use "Upload a voice" to send a file directly.

Everything else about a character (the identity lock, looks, the acting profile) is in the Casting Handbook; to put them in a story, see the Short Film Shoot Handbook or the Short Drama Shoot Handbook.