In Muse Video the soundtrack comes out of the same run as the frames, not from a second pass. That one fact decides everything else on this page: why ambience is nearly free, why a door closing is harder, why speech is the hardest thing in the medium, and why there is no such thing as fixing the audio.
A soundtrack is not one thing, and the three kinds of sound in it are not equally hard to generate. The ordering is not about how good they sound; it is about how precisely each one has to line up with the picture.
Continuous, formless sound with nothing to line up against. It can sit anywhere in the clip and still be right, which makes it the layer most likely to come back usable on the first attempt.
Each one is an event that has to land on the frame where the event happens. Close enough is not close enough: a sound a fraction late reads as wrong even when both halves are individually convincing.
Not one alignment problem but three stacked: mouth shapes against sounds, a voice that stays the same voice, and delivery that suits the picture. Why it is the hard case.
Look at the prompt library and you will notice the reliable entries request quiet room tone, wind, rain or no audio at all. That is not timidity it is choosing the layer that has nothing to get wrong.
Ask for the simplest soundtrack the shot can carry, exactly as you ask for the shortest clip. Both instructions reduce the number of things that must hold together, and both make the keeper arrive in fewer attempts.
The reason clips are short and fail late is temporal coherence: every frame has to agree with the frames before it. Audio adds a second axis to the same problem. Now the picture has to agree with itself, the sound has to agree with itself, and the two have to agree with each other at every instant.
Two convincing halves can still make an unconvincing clip. Sync is a third thing that has to be right, and it fails independently of both.
This is why a generated soundtrack tends to be either obviously fine or obviously wrong, with little in between. Continuous sound cannot really be out of step, so ambience passes. An event sound is either on the frame or it is not, and human hearing is unforgiving about which we notice tens of milliseconds of mismatch without being able to say what we noticed.
It also explains the absence that surprises people most: there is no audio-only retry. The soundtrack is not a stem the model attached at the end, so there is nothing to replace. A clip with good pictures and bad sound has to be generated again in full, and the new picture will differ in ways you did not ask for.
Talking heads are the thing people most want from generative video and the thing it is furthest from doing invisibly. Not one problem three, each sufficient to spoil a clip on its own.
Speech is the tightest alignment in the medium: not one event landing on one frame, but a continuous stream of them. We read lips more than we realise, and a mismatch of a few frames registers as uncanny even when nobody can name the fault.
Practical consequence: the more of a face is in frame and the closer it is, the more precision is demanded. Distance, profile angles and off-screen speakers all lower the bar.
Vocal identity is to audio what a face is to picture something that must persist across the clip. And like visual drift, the risk grows with every second, because there is more history to stay consistent with.
Practical consequence: short lines survive better than long ones, for exactly the reason short clips survive better than long ones.
Even with perfect lips and a stable voice, the performance can be wrong: too bright for the light, too warm for the room, an accent that does not belong to the place in the shot. This is the layer no technical fix addresses.
Practical consequence: if the words carry the message, put them in your own voice over a silent clip. The picture is what you came to a video model for.
What Muse Video can actually do with speech which languages, accents and voices, and how well it aligns is Meta’s to document and has not been published. Nothing here is a measurement; it is what the problem demands of any system attempting it. Treat Muse Voice Transcribe as a separate matter: transcription is the reverse direction.
Do not attempt to make a generated voice sound like a specific real person. Synthetic speech in someone’s likeness is the most consequential output any of these tools produce, it carries the heaviest disclosure obligations, and it is squarely against Meta’s terms and the law in many places.
Generic voices for generic characters are a different matter and worth labelling as synthetic wherever your clip lands, regardless of what the platform requires.
Watching and listening at once is how bad clips get approved: a strong picture covers for weak sound, and interesting sound distracts from drift. Separate the senses, then put them back together.
Turn away from the screen and play it. You are checking whether the sound is plausible on its own right room, right materials, no artefacts, nothing that wanders. Faults are obvious the moment the picture stops excusing them.
Now judge the picture the way you would any clip, with attention on the final second where drift appears. Two clean halves are the precondition for asking about sync at all.
Third pass, one question: does each sound land on its event? Footsteps, impacts, a cup meeting a table, lips against words. You are not re-judging quality here, only timing.
Most social video plays muted by default, and much of the rest plays through a phone speaker. If it only works on good headphones, it does not work. Social destinations.
If only the audio is wrong, remember there is no audio-only retry so weigh a full regeneration against silencing the clip in an editor and laying your own sound over it. Often the second is cheaper. What a run costs.
A Mac earns its place at exactly this step: a proper screen and a quiet room make the third pass possible in a way a phone does not why review is the desktop advantage.
For each one: whether to ask the model, and what to do if the answer is no. The pattern is that the model is good at atmosphere and you are better at meaning.
| What you want | Ask the model? | Better route |
|---|---|---|
| Atmosphere a room, weather, a street | Yes. This is the layer it handles most reliably. | Name the space and the materials in the audio clause, not just “ambient sound”. |
| A specific sound on a specific action | Worth one attempt, expect sync misses. | Ask for the action short and single, or add the effect yourself in an editor. |
| A voiceover carrying your message | No. | Ask for no audio, then record your own or use dictation to draft and read the script. |
| Dialogue between visible characters | The hardest thing here; treat as experimental. | Keep speakers off-screen or distant, or cut to reaction shots. Why it is hard. |
| Music under the clip | Not what this is for, and rights are murky. | Licensed music in your editor, where you also control the mix and the cut points. |
| Silence, to build the sound yourself | Yes say it explicitly. | “No audio” is a real instruction; leaving the clause out means inheriting a choice. |
Whether you can download the audio separately, or get stems, is unpublished assume a single file with the sound baked in and check what your own download contains before building an edit around it. What has and has not been announced.
Yes. Audio is part of what the model produces rather than something added afterwards, which is the single most important thing to know about it: the sound and the picture come out of the same run and cannot be regenerated separately.
No. There is no audio-only retry, because the soundtrack is not a separate layer the model attaches. If the sound is wrong, fixing it means generating the whole clip again and the picture will differ too. Why that follows.
Because ambience has nothing to line up with. Room tone, wind or traffic can sit anywhere in the clip and still be correct, while a footstep or a door closing has to land on the exact frame where the event happens the three layers.
It stacks three hard problems at once: mouth shapes have to match sounds frame by frame, the voice has to stay the same voice for the length of the clip, and the delivery has to suit the picture. Any one of them failing reads as wrong immediately.
Not published, and language coverage for audio rarely matches the countries where a feature is switched on. Test the language you need early in a project rather than at the end availability is layered.
Nothing has been announced about stems or separate audio export. Assume you get a single file with the sound baked in, and check what your own download actually contains before planning an edit around it.
Often, yes. If you intend to lay your own voiceover or music over the clip, asking for no audio avoids fighting a soundtrack you did not want and an unstated audio clause means inheriting whatever the model chooses.
Listen once without watching, watch once muted, then watch them together with attention on the last second. Faults in sound and picture hide each other when you take them in at the same time the three passes.
It matters less than the picture, which is an argument for asking for silence and adding your own sound rather than an argument for ignoring it. Something will play for the viewers who unmute, and if you never listened, you do not know what.
Do not attempt it. Imitating a real person’s voice is the clearest way to land on the wrong side of Meta’s terms and of the law in many places, and synthetic speech carries the heaviest disclosure obligations of anything these tools produce.
Room tone, weather and traffic come back clean because they have nothing to line up with. Words, timing and music are yours and a clip generated silent is the easiest one to finish in an editor.
Independent guide not affiliated with Meta. Languages, voices, sync accuracy and export behaviour are Meta’s to state; confirm them in Meta’s own documentation before planning around them.