By Mirelo Team |
Add sound effects to a video while keeping its dialogue in three steps:
The default result is a finished mix with dialogue, effects, and ambience. Separate dialogue and effects-only outputs need the API options described below. Keep your source recording: speech preservation does not guarantee that the voice sounds identical or that its waveform is unchanged.
Read the SFX 2.0 API guide or explore Mirelo SFX.
A video with good dialogue and weak sound still feels unfinished. This often happens with AI-generated video: the voice works, but footsteps, doors, traffic, and room sound are thin, generic, or missing. Replacing the entire soundtrack would also discard the voice.
Speech preservation lets you keep the voice from the source clip while generating new sound for the scene. It works with supported clips from AI video generators, avatar tools, ads, interviews, tutorials, and vlogs. Developers can configure it in the API.
Mirelo SFX 2.0 generates sound effects from video. It uses the footage to place sounds on visible actions, such as a footstep, a door closing, or a splash. You do not need to describe and place every sound on a timeline; the video provides the context.
Speech preservation brings the same video-driven generation to clips that already have dialogue.
mix with dialogue preserved by default. This is the normal result, not an automatic set of separate tracks.controls.separate_speech: true to also receive a speech output when speech is detected. This is dialogue separated from the input soundtrack.controls.preserve_speech: false and controls.separate_speech: true. The result contains an effects-only mix, the same effects-only audio as sfx, and the separated speech when detected.With speech preservation on, the generated mix already contains dialogue as the model renders it. Do not layer the separate speech output on top of that mix, or you will double the dialogue. If exact voice fidelity matters, use your original dialogue recording with the effects-only output.
Review the result for changed words, voice quality, timing, and separation artifacts. Download any separate speech file before its link expires; it is not stored as a generation you can retrieve again.
AI video creators. Add sound to a generated clip that already has dialogue.
AI video platforms. Offer an audio step after video generation. Send a finished clip to the Mirelo API and return the mix or the separate tracks to your users.
Avatar, ad, and explainer platforms. Add product sounds, room tone, and ambience around spoken content.
Editors, agencies, and studios. Build sound design around a recorded interview or voiceover. Keep the source recording when an exact, approved vocal track is required.
Upload the clip to the API playground, select SFX 2.0, and leave Preserve speech enabled. The video guides the sound and timing. Listen to the result before using it in your edit.
Use SFX 2.0 with Preserve speech enabled. It uses the existing dialogue while generating effects and ambience around it. The default output is one mix; request separate speech explicitly if you need it.
SFX 2.0 can create new background sound around your dialogue, but the voice may sound different. To keep the original voice exactly, combine your original dialogue recording with the effects-only output.
In the API, request controls.separate_speech: true. When speech is detected, the result includes a separated dialogue track. Separation may leave artifacts, including traces of background sound; it is not a guarantee of perfect music or noise removal.
Yes. Request controls.preserve_speech: false with controls.separate_speech: true. Use the effects-only output and separated dialogue in your editor, or use your original dialogue recording if you need to preserve it exactly.
SFX 2.0 uses the input dialogue, but the generated mix renders the voice as part of its output. Words, tone, or timing may change. The optional speech output is separated from the input instead; separation can also affect its sound.
Review the balance in the finished mix, especially when source dialogue is quiet or noisy. If the result masks words, use effects-only output and mix it with your dialogue at the level you need.
Yes. Add a short prompt or negative prompt in the API to guide the result. On SFX 2.0, these are guidance rather than exact instructions; the video still drives generation. Compare the returned versions before selecting one.
SFX 2.0 accepts a silent clip and generates its sound. If no speech is detected, a request for separate speech returns no speech output. If your dialogue is in another file, send it as input.audio aligned with the video. Otherwise, no separate audio input is needed.
SFX 2.0 supports up to 10 minutes of video in Studio and the API, with or without speech preservation. For longer edits, contact Mirelo about Enterprise. Developers can read the current operations.video-to-sfx.duration_ms limits from GET /v3/models/sfx-2.0 before sending a job.
Add Mirelo as an audio step after your video model. Send each finished clip to the API with Preserve speech enabled when it has a soundtrack you want to retain. Read the SFX 2.0 integration guide.
Mirelo uses the finished video as input rather than requiring a particular video model. The file still needs to meet the API's supported input and duration limits.
Yes. Offer effects-only generation and optional speech separation when users need to mix tracks themselves. With speech preservation enabled, the normal result is a mix; it does not automatically include an effects-only track.
SFX 2.0 is priced per second of generated audio, per variant. Preserving speech or requesting the separate speech output adds no extra charge on SFX 2.0. Use API preflight for the current quote and check the completed job for the settled charge. See usage rates.
Yes. Jobs run in the background and can run in parallel within the limits listed in the API docs.
It means using dialogue from an existing soundtrack while creating new sound around it. On SFX 2.0, the generated mix can change the voice. Speech separation is the distinct option for returning dialogue isolated from the input.
Text-to-SFX creates a sound from a written description that you place in the edit. Video-to-SFX uses the footage to generate sound timed to visible action. Mirelo offers both.
See the SFX 2.0 guide and video-to-SFX reference in the Mirelo API documentation.