
By Mirelo Team |
The Mirelo Sound Agent uses AudioShake's separation models to take apart the sound that is already in your video, so you can keep the parts you want and change the rest by asking. Ask in plain language and AudioShake removes the music while keeping the dialogue and effects, splits the dialogue from the music-and-effects bed, puts each speaker on their own track or cleans noise and echo out of a voice. The Sound Agent then rebuilds the sound around what you kept, with new sound effects and ambience from Mirelo SFX 2.0, new music or a dub in another language, all in the same project.
Most videos arrive with their sound already mixed into one track: a generated clip with music nobody prompted for, an interview recorded next to a humming air conditioner, a finished cut delivered without its project files. Adding new sounds only helps once you can get at the parts worth keeping, which is the job AudioShake does inside the Sound Agent. AudioShake builds AI that separates audio into stems from any recording, and whenever a request needs the existing audio taken apart, the agent sends it to AudioShake.
The agent calls AudioShake's state-of-the-art models for four kinds of requests and picks the right one based on your request.
| What you ask for | AudioShake model | What lands in your project |
|---|---|---|
| Take the music out | Music Removal | The dialogue and the effects, with the music removed |
| Split the dialogue from everything else | Dialogue, Music & Effects | A dialogue track and a music-and-effects bed |
| Give each speaker their own track | Multi-Speaker Separation | One track per speaker |
| Clean up a voice | Speech Denoise and Speech DeReverb | One version with the noise removed and one with the echo removed |
Everything around those steps runs in the Sound Agent itself: it analyses the picture and the sound, generates sound effects and ambience with Mirelo SFX 2.0, generates music, dubs dialogue and places each part on the timeline.
AI-generated clips often come back with thin effects or room tone that shifts from one shot to the next, while the dialogue is the one part you want to keep as recorded. Mirelo SFX 2.0 can keep dialogue on its own, using the input dialogue while it generates new effects and ambience. When the exact voice matters, ask the agent to separate the dialogue first: AudioShake splits the original dialogue from the music-and-effects bed, and the agent keeps that dialogue on its own track and generates sound without the original dialogue, so the speech in your final mix comes from the original recording.
Try asking: "Keep the dialogue and replace the sound effects with rain and city traffic"
Music that came baked into a generated clip, a track you cannot clear, or a score the client changed their mind about can all come out without losing the rest of the soundtrack. AudioShake's Music Removal model takes the music out and keeps the dialogue and the effects, so footsteps, doors and room tone stay where they were, and from there you can leave the scene without music or ask the agent for a new track in a different direction.
Try asking: "Remove the background music and keep the dialogue and footsteps"
Taking a video into another language usually means rebuilding its music and effects as well, unless you can lift the original speech out of the mix. AudioShake splits the dialogue from the music-and-effects bed, the agent dubs the dialogue into the language you ask for, and the new voice sits on top of the original music and effects.
Try asking: "Dub this into Spanish and keep the original music"
Podcasts, panels and interviews put several voices on one track, which makes it hard to bring up a quiet guest or cut a moment where two people talk over each other. AudioShake built Multi-Speaker Separation to give each speaker their own stem, even when the speech overlaps, and the agent puts every voice on its own track so you can turn one up or cut one without affecting the others.
Try asking: "Split this interview into one track per speaker"
Remote interviews, voiceovers recorded at home and dialogue shot in a hard-walled room tend to pick up hum, background noise or echo. AudioShake's Speech Denoise takes out background noise and hum while keeping the natural sound of the room, and Speech DeReverb also removes the room's reflections and echo for a drier voice. The agent gives you one cleaned version from each, so you can compare the two and keep the one that fixes your recording.
Try asking: "Remove the echo from my voiceover"
A finished video without its project files leaves you with one mixed track, so a voice buried under the music cannot simply be turned up. AudioShake splits that track into the dialogue and a music-and-effects bed, and the agent puts both on the timeline, where you can bring the voice up, turn the bed down or keep working on either one with the next prompt.
Try asking: "Split the dialogue from the music and effects, then make the voice louder"
The Sound Agent works out which separation your request needs, and when a request could mean two things, such as "split the stems" right after it has generated sound effects, it asks one short question before it runs anything. Separation costs credits, and the agent asks first when a step goes over your spending limit.
AudioShake then separates the audio, and each part comes back as its own track on the timeline. If a part turns out to be silent, such as a dialogue track from a clip with no speech, the agent leaves it out and says so in one line. From there you carry on with the next prompt, whether that is new sound around what you kept, a change in levels, a dub or an undo.
Both options are in the Sound Agent, and the choice depends on how close the dialogue has to stay to your original recording.
| SFX 2.0 keeping the dialogue | Dialogue separated by AudioShake | |
|---|---|---|
| What happens to the speech | SFX 2.0 uses the input dialogue while it generates new effects and ambience, and words, voice quality or timing may change | AudioShake separates the original dialogue, which stays on its own track while new effects-only sound goes underneath |
| Steps | One generation | A separation, then a generation |
| Best for | A fast sound pass where close to the original is enough | Interviews, ads, dubs and any scene where the exact voice matters |
For everything else the agent can do, read Meet the Sound Agent, and for keeping speech while changing the rest of the mix, see Add AI sound effects without losing dialogue.
Upload the video to Mirelo Editor and ask the Sound Agent to remove the music. It runs AudioShake's Music Removal model, which takes the music out and keeps the dialogue and the effects, and you can then ask for a new track or leave the scene without one.
Yes. Ask the Sound Agent to separate the dialogue first: AudioShake puts the original dialogue on its own track, and the agent generates effects-only sound to go underneath it. SFX 2.0 can also keep dialogue in a single generation, though words, voice quality or timing may change.
AudioShake runs the Sound Agent's separation: removing music, splitting dialogue from music and effects, separating speakers, and cleaning noise and echo out of speech. The agent handles the work around it, including analysing the video, generating sound effects with Mirelo SFX 2.0, dubbing and editing the timeline.
Yes. AudioShake built Multi-Speaker Separation to split each speaker into their own stem even when their speech overlaps, and the Sound Agent puts each one on its own track.
Close, but not always identical. AudioShake's models are built to keep the voice intact and leave as few artifacts as possible. Listen to each separated track before you export, and undo the step if a part doesn't sound right.