Seed Audio 1.5: translate and dub video in the same voice
One model for voice, sound effects and music, driven by your video
Seed Audio 1.5 is ByteDance's new audio model. It takes a video, listens to it and writes a new soundtrack: the same narrator speaking Japanese, a dub that follows the picture, or the voice, music and effects pulled apart into separate tracks. We used it in early access to translate a 9-minute animated film into 12 languages. Here is what it did, where it slipped, and how to get the same result.
Doitong is an independent platform and is not affiliated with ByteDance.
Our test: one film, 12 languages, the same narrator
We took “The Last Ordinary Day in Pompeii”, a 9-minute animated history film with English narration, and ran it through Seed Audio 1.5 video translation once per language. Play the video and open Settings → Audio track to switch between the English original and the 12 translations.
Audio tracks on this video: English (original) plus Russian, German, French, Spanish (Spain), Italian, Japanese, Korean, Dutch, Swedish, Danish, Norwegian and Finnish.
- Film length
- 9:00
- Languages
- 12
- Time per request
- 2.6–3.3 min
- Length difference
- < 0.03 s
How we ran the test
The numbers above come from our own logs, not from a press release. Three details matter if you want to repeat it.
Split at a pause
One request accepts a video of 4 to 360 seconds, so the 9-minute film went in as two parts cut at a natural pause: 4:43 and 4:17. Twelve languages meant 24 requests, and each one came back in 157 to 200 seconds.
The length held
Every translated part came back the length of the original. Part one returned at 283.26 seconds against 283.29 for the source, part two at 256.83 against 256.83, and the figures were the same in every language. The new narration dropped onto the picture without retiming.
We sent the voice alone
We fed the model the narration without music and mixed our own score and effects back afterwards. Keeping music and effects inside the translation is a stated feature of the model, but this particular test did not rely on it.
What Seed Audio 1.5 can do
Version 1.5 is the first Seed Audio model that accepts video. That one change opens up most of the jobs below.
Translate a video in the same voice
From English or Chinese into 30 languages. The voice keeps the speaker's character, and the original effects and background music stay in place while only the speech changes.
Dub a video from a description
Give it a video and describe what should be heard: the lines, the voices, the effects, the music. It writes audio that follows the picture. It will not invent the script for you.
Split audio into stems
Feed it an audio file and get voices, effects and music back as separate tracks, up to six of them, ready for a mixing session.
Score a silent clip
Ask for ambience, footsteps, a door, a music bed, and place each one on the timeline with timestamps in the prompt.
Clone a voice
Up to six reference clips of 2 to 30 seconds each set the timbre. Combine a reference with a video and the dub is spoken in that voice.
Write songs with lyrics
Music up to six minutes long from a text brief, with sung lyrics in English, Chinese, Japanese, Korean or Cantonese.
Seed Audio 1.5 limits
Limits from ByteDance's early access integration guide. Check them before you cut your video.
| Input | Limit |
|---|---|
| Video length | 4–360 seconds per request |
| Video formats | MP4, MOV |
| Video resolution | 480p, 720p or 1080p |
| Video file size | Up to 200 MB |
| Voice reference clips | Up to 6, 2–30 s each, 180 s in total |
| Audio for stem separation | One file, 2–360 s, up to 90 MB, WAV or MP3 |
| Output | Value |
|---|---|
| Longest audio per request | 6 minutes |
| Stem tracks | Up to 6 |
| Formats | WAV (default), MP3, PCM, OGG Opus |
| Default sample rate | 44.1 kHz |
| Languages for speech | 30 |
| Video translation source | English or Chinese, or detected automatically |
Source: ByteDance Seed Audio 1.5 early access integration guide, October 2026. Limits may change after release.
30 languages
Seed Audio 1.5 speaks these languages. Video translation goes from English or Chinese into any of them.
Lyrics in songs are supported in a smaller set: English, Chinese, Japanese, Korean and Cantonese.
Prompts that work
ByteDance's prompt guide describes three levels: a plain description, a script that sets the order of events, and a timeline with exact seconds. Start plain and add timestamps only where timing matters.
A plain voice-over
Generate only this speaker as dry, close-miked speech. No music, no sound effects, no room reverb, no second speaker. A calm adult female voice at a steady pace reads: “Most people in Pompeii woke up that morning expecting an ordinary day.”
The constraints first, then the voice, then the exact words. Saying what must not appear matters as much as what must.
A timed scene
[0.0s:12.0s] Soft market ambience under the whole scene. [0.5s:1.5s] A wooden cart rolls past. [2.0s:6.5s] An old man, warm and unhurried, says: “Bread, still warm from the oven.” [7.0s:12.0s] A low drum rumble rises in the distance.
Timestamps place each sound on the timeline, so the audio lines up with cuts you have already made.
What to watch out for
Things we hit in testing, and things the documentation is clear about.
- Names and rare words can drift between takes. In the Russian translation of our second test film, set in London in 1666, the old word for a Thames river boat came out wrong and one character's name changed form. Video translation accepts a glossary, so list the names before you run it.
- Dubbing needs a description. Seed Audio 1.5 does not write lines on its own; tell it what is said and what is heard.
- It makes sound, not lip movement. A translated voice keeps the timing of the original, but the mouths on screen are not redrawn.
- One request covers up to 6 minutes. Longer videos have to be split, and a cut in the middle of a sentence is audible, so cut at a pause.
- Video translation starts from English or Chinese. A video in another language cannot be translated directly.
- Clone only voices you have the right to use.
Seed Audio 1.5 FAQ
What is Seed Audio 1.5?
Seed Audio 1.5 is an audio generation model from ByteDance, released on October 28, 2026. It produces speech, sound effects and music in one pass and is the first Seed Audio version that accepts video, which is what makes video dubbing and same-voice video translation possible.
Does Seed Audio 1.5 translate a video in the same voice?
Yes. Video translation re-voices the speech in another language while keeping the speaker's voice character, and keeps the original sound effects and music. In our test a 9-minute English narration came back in 12 languages, each the same length as the original.
Which languages does Seed Audio 1.5 support?
30 languages for speech, including English, Chinese, Japanese, Korean, Spanish (Spain and Mexico), Portuguese (Brazil and Portugal), German, French, Russian, Arabic and Hindi. Video translation goes from English or Chinese into any of the 30.
How long can the video be?
One request takes a video of 4 to 360 seconds, up to 200 MB, in 480p, 720p or 1080p. For a longer film, split it at pauses and translate the parts one by one, as we did with the 9-minute Pompeii film.
How long does a translation take?
In our runs each part of about four and a half minutes took between 157 and 200 seconds per language.
Does Seed Audio 1.5 do lip sync?
No. It generates audio. The translated voice follows the timing of the original, but the video itself, including mouth movement, is not changed.
Can it separate vocals, music and sound effects?
Yes. Give it an audio file of up to 6 minutes and it returns voices, effects and music as separate tracks, up to six in total.
What changed from Seed Audio 1.0?
Seed Audio 1.5 adds video input, which is what brings video dubbing and video translation. It also adds stem separation and music up to 6 minutes, accepts up to six voice reference clips, produces up to 6 minutes of audio per request and speaks 30 languages.
Is Seed Audio 1.5 the same as Seed Audio 2.0?
Some sites list the upcoming model as Seed Audio 2.0. ByteDance's early access documentation calls it Seed Audio 1.5, and that is the name we use.
Is Doitong affiliated with ByteDance?
No. Doitong is an independent AI video platform. We had early access to Seed Audio 1.5 and describe our own results here.
What you can do with Seed Audio 1.5
Give your video a voice in another language
Start with one short clip, check the names, then run the whole film.
Start in Doitong