Mirelo SFX 2.0 removes unwanted music while preserving speech

• Unwanted music can be replaced without rerendering the picture, as AIwithSynthia's comparison demonstrates with Mirelo SFX 2.0.
• Speech preservation requires existing dialogue. The model does not create new spoken lines, according to venturetwins' clarification.
• Completed edits can receive a coherent sound pass for up to 10 minutes, as aakashgupta reported from the launch.

Mirelo's launch FAQ says it regenerates the existing spoken performance with new effects around it. The fal API defaults to two soundtrack variations, and Mirelo's direct API pricing adds a separate speech-preservation charge.

AIwithSynthia described the choice between muting every sound and gambling credits on another video generation, then said Mirelo's approach worked for them.

Mirelo describes three steps:

1. Detect and preserve the existing dialogue.
2. Remove existing non-dialogue audio, including unwanted music.
3. Generate fresh effects, foley and ambience synchronized to the picture and fitted around speech.

The original effects can therefore be replaced along with the music. Keeping the picture untouched is a welcome escape from rerolling a keeper shot.

AIwithSynthia used a battlefield clip to demonstrate another repair: adding battle sounds where the original generation leaned on dramatic music.

Existing voices remain part of the input. The model supplies the surrounding soundscape, with speech retained in the comparison shared by venturetwins.

For generating dialogue, venturetwins outlined a three-tool workflow:

1. Create speech with ElevenLabs.
2. Feed that audio into an audio-conditioned video model such as Seedance.
3. Send the resulting video to Mirelo to build the broader soundscape while preserving speech.

Seedance can still add audio nobody requested, venturetwins noted in a follow-up, even when its handling of the supplied speech works well.

Separately generated shots can each invent their own background noise. Writing about continuity, aakashgupta used the example of five clips of the same kitchen arriving with five subtly different acoustic backgrounds.

Mirelo says SFX 2.0 maintains effects and ambience across cuts, allowing an assembled sequence to receive one sound pass. Its announced duration limits are:

• Mirelo Studio and direct API: up to 10 minutes.
• Enterprise: unlimited duration available upon request.

Mirelo's published comparisons are company-run blind listening tests. For each comparison, 25 raters judged 30 videos on audio quality and fit with on-screen events, with a tie option.

Two results against native video-model audio:

| Original audio | Mirelo preferred | Original preferred | Tie |
| --- | ---: | ---: | ---: |
| Veo 3.1 Fast | 57% | 21% | 22% |
| Seedance 2.5 | 47% | 28% | 25% |

The launch's claim that Mirelo sounded “as good as or better” than Seedance in 72% of cases combines 47% wins with 25% ties.

AIwithSynthia urged readers to test a live demo alongside the benchmark numbers.

Creator QualitativAi described a separation-based alternative for retaining the original effects:

1. Use Moises AI to isolate music, effects and dialogue into separate tracks.
2. Delete the music.
3. Recombine the remaining audio, or keep effects and dialogue separate in the editor.

QualitativAi said they automate this with Astra, keeping the video and associated audio files together for timeline use. Dialogue sometimes still needs EQ or cleanup, particularly when the original music is heavy.

The fal endpoint documentation exposes controls beyond uploading a video:

• preserve_speech: defaults to true. Setting it to false generates effects only and suppresses speech.
• text_prompt: optional guidance for the generated soundtrack.
• negative_prompt: a short, comma-separated list of sounds to exclude.
• duration: defaults to 10 seconds. Longer outputs are generated in 10-second windows.
• num_samples: defaults to two variations.

The endpoint returns the input video with each generated soundtrack as an MP4.

Mirelo's launch announcement lists these access points:

• Mirelo Playground and Editor.
• Mirelo API and MCP.
• Runway and fal.
• Premiere Pro and DaVinci Resolve plugins, plus a Roblox plugin marked coming soon.

On fal, introductory pricing is $0.010 per output second, per sample, from October 14 through December 13, 2026. The standard $0.015 rate resumes December 14; two 10-second variations cost $0.20 during the promotion and $0.30 afterward.

Mirelo's own usage table lists different billing:

• Generation: 10 credits per output second, per variant, discounted from 15 through December 12, 2026.
• Preserve Speech: an additional 8 credits per second of source audio when speech is detected, charged once rather than per variant.

There is also a surface discrepancy: the launch promotes speech-preserving workflows broadly, while that usage table marks Preserve Speech as “Not available” in Studio and available through the direct API.