One click, whole workflow
Recipes are here. Package a whole workflow and run it in one click
Explore recipesBest AI Music and Sound Effect Generators (2026)
The best AI sound effect generator is ElevenLabs SFX in Fuser and Lyria 3.5 is best for songs, two of 9 audio models on the canvas, with MMAudio and F5 TTS compared by job.
FuserUpdated
All guides · Best image-to-video models
Quick answer: use ElevenLabs SFX in Fuser for sound effects from a text prompt and Lyria 3.5 in Fuser for a full song or an instrumental from one prompt. Fuser runs 9 audio models on the same canvas as its 128 image and 162 video models, so the sound pass happens right next to the picture. Add MMAudio or Mirelo SFX for Foley that follows the action in a silent AI video, ElevenLabs TTS for expressive voiceover and F5 TTS to clone a voice from a short sample. Use MiniMax Music 3 when a song or cue has to fit a set length, and MiniMax Music to write a song over the style of a reference track. Most AI video still arrives silent, so the sound pass is where a clip starts to feel finished.
Sound effects from a prompt: ElevenLabs SFX
ElevenLabs sound effects turn a description into Foley, impacts, UI cues, risers and room tone. In Fuser you can set a duration from 0.5 to 22 seconds. Describe the material, the action and the space. The effect on the canvas above came from one sentence: a ceramic vase set down on a stone plinth, a soft hollow clink, then a faint ring. It is the wrong tool for full songs or intelligible dialogue.
Sound for silent video: MMAudio and Mirelo
When you already have the clip, let the model watch it. MMAudio analyses the motion and generates synchronised effects and ambience; in Fuser it returns the video with the new track, from 1 to 30 seconds. For longer shots, split the clip into short segments, because sync drifts on long inputs. It also has a text-to-audio mode when there is no video. Mirelo SFX is built for the same job on commercial cuts, trailers and social video, handles up to 60 seconds and accepts an optional prompt to steer the sound. Neither generates music or speech.
Voiceover: ElevenLabs v3 and F5-TTS
ElevenLabs v3 is the choice for expressive narration and character dialogue. Audio tags such as [calm] or [short pause] direct the performance inside the script, as in the voiceover above. Split long scripts into sections to keep the voice consistent.
F5-TTS clones a voice from a clean reference clip of around ten seconds, including across languages. Adding the reference transcript speeds up generation. Keep each line under about 30 seconds, and do not expect it to invent crying or shouting that is not in the sample.
Songs and score from a prompt: Lyria 3.5 and MiniMax Music 3
Google lists Lyria 3.5 for "full-length songs with verses, choruses, bridges" lasting "a couple of minutes", and says every generated track carries a SynthID audio watermark. Fuser's Lyria 3.5 node takes one prompt and an optional image for inspiration: style, tempo, lyrics and length all go in the prompt, and it returns an MP3 plus a Lyrics output with the song's sections. Length is only a hint. In our Lyria 3.5 vs MiniMax Music 3 test, "a 60-second track" came back at 66.0 and 60.3 seconds, and "a 30-second track" at about 61 seconds both times.
MiniMax introduced Music 3.0 as an open-weights model for complete songs of up to five minutes. Fuser's MiniMax Music 3 node takes a style prompt and a separate, required Lyrics field, with section tags such as verse] and [chorus] on their own lines, plus Duration, Seed, Steps and Guidance. Duration is a ceiling: in the same test no run went more than a tenth of a second past it, but four of six stopped early, so set it a few seconds longer than the slot and trim. Scoring a 10-second clip with Duration set to 10 gave us a track of exactly 10.00 seconds ([AI music for video). For an instrumental, our MiniMax Music 3 prompt guide got a clean result from a Lyrics field of section tags alone ([intro], [instrumental], [solo], [outro]), while [instrumental] on its own produced wordless vocals in one run.
Fuser charges Lyria 3.5 one flat amount per song and MiniMax Music 3 per second of the Duration you set, so MiniMax is the cheaper of the two below about 50 seconds and the dearer above. Choose Lyria 3.5 for full songs, instrumental beds and music that takes its mood from a picture. Choose MiniMax Music 3 for a cue that must fit a fixed slot, or when you want a seed you can reuse.
Songs from a reference track: MiniMax Music
Fuser's MiniMax Music node takes lyrics (up to 600 characters) and a reference song longer than 15 seconds with vocals and music, and returns a new song in that style. It suits demos, genre flips and campaign jingles. It is not a text-only music generator and is not built for instrumentals, so bring a reference track you have the right to use. For songs described in words, use Lyria 3.5 or MiniMax Music 3; MiniMax Music 3 vs Music-01 compares the two MiniMax nodes on the same lyrics.
Put sound in the same workflow as the picture
In Fuser, keep the audio nodes next to the image and video steps they belong to. Generate a still, animate it, then send the clip to MMAudio or Mirelo and your script to ElevenLabs, so a new cut picks up fresh sound without re-exporting anything. A music node returns its own audio file, which you join to the picture in your video editor, or give it to Wan 3.0 as reference audio to build a new shot around the track; AI music for video walks through both. See the full multi-model workflow, and save the graph as a Recipe to reuse it for every edit.
Last verified October 5, 2026.
Choose the audio model for the job.
Match what you are starting from to the right node.
| Model | Reach for it when | Watch out for |
|---|---|---|
| Effects and Foley | ||
| ElevenLabs SFX | Effects from a text prompt, 0.5 to 22 seconds. | Not for songs or speech. |
| MMAudio | Synced effects for a silent clip, 1 to 30 seconds. | Sync drifts on long clips; no music or speech. |
| Mirelo SFX | Foley for commercial cuts and social video, up to 60 seconds. | Effects only; no score or dialogue. |
| Voice | ||
| ElevenLabs TTS (v3) | Expressive narration directed with audio tags. | Split long scripts to avoid drift. |
| F5 TTS | Cloning a voice from a clean ~10-second sample. | Keep lines under ~30 seconds. |
| Music | ||
| Lyria 3.5 | Full songs or instrumentals from one prompt; optional image for mood; returns lyrics too. | Length is a prompt hint; our 30-second requests came back at about a minute. |
| MiniMax Music 3 | Songs from your lyrics and a style prompt, with a Duration ceiling of 1 to 300 seconds. | Often stops before the ceiling; the Lyrics field cannot be empty. |
| MiniMax Music | Songs from lyrics plus a 15-second-plus reference track. | Needs a reference with vocals; not for instrumentals. |
Questions, answered.
ElevenLabs SFX in Fuser is the strongest choice when you start from a text description. If you already have a silent video, run MMAudio or Mirelo SFX in Fuser to generate effects that follow the action on screen, beside the image and video nodes that made the clip.
Send the clip to a video-to-audio model such as MMAudio or Mirelo SFX. They analyse the motion and return the video with a synchronised effects track. Split long clips into short segments for tighter sync.
Yes. Lyria 3.5 writes a full song, with your lyrics or its own, from one text prompt, and MiniMax Music 3 sings lyrics you supply in a style you describe, up to five minutes. Fuser's original MiniMax Music node works differently: it copies the style of a reference track longer than 15 seconds.
MiniMax Music 3. Its Duration setting is a ceiling: in our runs it never went more than a tenth of a second past it, though it often finished early, so set it slightly longer than you need. Lyria 3.5 takes length only as a hint in the prompt; our 30-second requests came back at about a minute.
ElevenLabs v3 is built for expressive, directed performances using audio tags. F5-TTS is the better fit when you need to clone a specific voice from a short sample.
Yes. In Fuser, connect a script to a voice node, a clip to a video-to-audio node and a prompt to a sound effect node on the same canvas, then swap or rerun any of them.
Give every clip a finished sound.
Generate effects, voice and music next to your images and video, then reuse the workflow.
