Why does my character's voice change between scenes?

Because each generation is independent. Unless you send the same voice sample as a reference every time, the model invents a new voice that merely fits the description, and 'a gravelly man in his fifties' has thousands of valid readings. The fix is mechanical: attach one audio sample to the character and re-send it with every render.

How long can the reference audio be?

The cap is on the total length of all audio references in one request, not per clip: about 15 seconds for the Seedance 2.0 family and 30 seconds for 2.5. It does not scale with the length of the video you asked for — a 5-second render has the same audio budget as a 15-second one. Two characters with 10-second samples will never fit in one shot together.

What audio format does the API accept?

MP3 and WAV only, each clip at least 2 seconds and under 15 MB. An OGG voice note from a messaging app is refused. If you send a data URL, note that the provider reads the word after 'audio/' as a file format name rather than a MIME type: 'audio/mp3' is accepted and the technically correct web MIME type 'audio/mpeg' is rejected with an unhelpful 'Invalid base64 audio_url'.

Should the voice sample have background music?

No. The sample becomes the reference for every future scene, so anything in it gets copied forward — music, room tone, a second speaker. Use a dry recording of one voice. When generating a sample with a text-to-speech model, say so explicitly in the prompt, because such models will happily produce a whole sound scene.

A voice that survives the cut

Faces get all the attention in AI video, but the thing that actually breaks a series is sound. Your captain is a weathered baritone in scene one and a bright thirty-year-old in scene four, and no amount of visual consistency rescues it — the audience hears a different person. The fix isn't prompt craft. It's mechanical, and it comes with limits that no documentation states plainly.

Why the voice drifts

Every generation is independent. The model has no memory of the previous shot, so a described voice — "a gravelly man in his fifties" — is re-invented from scratch each time, and that description has thousands of legitimate readings. You are not asking for the same voice; you are asking for a voice from the same category, and getting one.

The only thing that pins a voice down is the voice itself, supplied as a reference on every single request. Send the same few seconds of audio with each render and the model copies the timbre instead of guessing at it.

The practical rule. One sample per character, attached to the character rather than to a scene, re-sent automatically with every generation that features them. Do this by hand and you will eventually forget on one shot — and that one shot is the one that ends up in the edit.

The limits nobody writes down

We measured these against the live API, because the published documentation doesn't state them in a usable form.

Constraint	Value	Consequence
Total audio per request	≈15s on the 2.0 family, ≈30s on 2.5	The cap is on the sum of all audio references, not per clip
Relation to video length	None	A 5-second render has the same audio budget as a 15-second one
Clip length	2s minimum, 15 MB maximum	Very short samples are refused outright
Formats	MP3 and WAV only	A voice note recorded in a messenger (OGG) has to be converted
Audio references per request	3 on the 2.0 family, 10 on 2.5	Ensemble scenes need 2.5, and even there the total-length cap binds first

The first row is the one that bites. Because the ceiling applies to the total, two characters with ten-second samples can never appear in the same shot on a 2.0 model — the request is refused before anything renders. Keep samples short: four to six seconds carries a timbre perfectly well and leaves room for a second and third speaker.

A format trap worth knowing. If you pass audio as a data URL, the provider reads the word after audio/ as a file format name, not as a MIME type — and it knows exactly two, mp3 and wav. So data:audio/mp3;base64,… works, while audio/mpeg — the technically correct web MIME type for an MP3, and the one every library will hand you — is rejected with Invalid base64 audio_url. The error blames your encoding, which is fine. Your encoding is not the problem. This one cost us an afternoon.

What makes a good sample

The sample is not a demo reel; it is the specification for every future scene. Everything in it propagates.

One voice, nothing else. No music, no effects, no room tone, no second speaker. Whatever shares the clip gets copied into scenes where you never asked for it.
Ordinary delivery. Record the register you want across the series, not a dramatic peak. A shouted sample makes a character who always sounds strained.
Four to six seconds. Enough for timbre, short enough to leave room for other speakers under the total cap.
If you generate it, say so explicitly. Text-to-speech models are perfectly happy to produce a full soundscape with music underneath. The instruction has to spell out one speaker, no music, no effects, no background — otherwise you have just made an orchestra a permanent member of your cast.

Generating a sample is cheap — roughly $0.002 per second of audio, so a four-second take runs about eight tenths of a cent. Iterating on it a dozen times costs less than one video render, which is the right ratio: the sample outlives every shot it appears in.

How Longtake handles it

The mechanism above is wired in rather than left to discipline. A character owns a portrait and a voice sample; writing @captain in a prompt attaches both, and the studio adds an explicit instruction tying the character's image to their audio reference so the model doesn't have to infer the pairing. The composed prompt is shown to you exactly as it was sent, so nothing about it is a black box.

The total-length ceiling is checked locally before the request goes out, so an over-budget shot fails instantly with a readable message instead of an API error about r2v limits. Sample duration is measured on upload and stored, which is what makes that check possible.

Questions

Can I use a real actor's voice?: Technically yes — a dry MP3 works regardless of origin. Whether you may is a rights question, not a technical one: get the performer's consent in writing, the same as you would for any recording you intend to reuse indefinitely.
Does the voice reference also control what the character says?: No. The line is in the prompt; the sample supplies timbre and delivery style. Keep dialogue in quotation marks in the prompt text.
Why is my two-character scene refused when each sample is fine alone?: Because the limit is on the total. Two ten-second samples make twenty seconds, over the 2.0 ceiling. Shorten both, or move that shot to 2.5.
Does a longer sample give a better match?: Not in our experience, and it costs you the budget for other speakers. Four to six clean seconds beat twelve noisy ones.

Why reference photos get rejected, and what passes → Getting a Seed Speech key for voice generation → How this compares with credit-based tools →

Measured in August 2026 against Seedance 2.0, 2.0 Fast and 2.5 through BytePlus ModelArk. Provider limits change; the reasoning outlives the numbers.