Mirelo SFX 2.0: sound effects and ambience from a sentence

Foley artists famously clap two coconut halves together to make horse hooves. Mirelo SFX 2.0 skips the coconuts: you describe the sound in words and it returns one audio file of sound effects or ambience. The example on its model page is a good test of what it can do, a single alien spacecraft flyby that starts as a distant rumble, swells into a turbine roar overhead, drops in pitch with a Doppler shift and fades away. That is a sound with a beginning, a middle and an end, written as one paragraph.

What can it make?

Two kinds of things. Events: a door slam, a sword drawn from a sheath, a flyby, glass breaking. Beds: rain on a tin roof, a busy café, wind in a forest, a spaceship engine room hum. The duration slider runs from a tenth of a second to ten minutes and starts at 30 seconds, so both a short click and a long loop for a scene fit in one model.

How should I write the prompt?

Like a sound designer's brief, not like a caption. Four things make the difference: the source (what makes the sound), the action (what happens, in order), the space (a small room, a cathedral, open air) and the distance (close up, across the street). The flyby example has all four, and that is why it works.

What is the negative prompt for?

Keeping out what sneaks in. Ask for a busy street and you may get a car horn you did not want, or a voice in the café. The negative prompt sits under the advanced settings and its own hint lists the usual suspects: music, voices, traffic. For ambience under dialogue, putting 'voices, speech, music' there keeps the bed clean so your actors are the only people talking. The negative prompts guide explains the idea in more depth for pictures and video; it works the same way here.

How do I get variations of the same sound?

Use the seed. Left empty, every run is a new random take, which is what you want while exploring. When you hear a take you like, set its seed and keep it while you adjust the wording: the result changes only as much as your words do. A game needs five footsteps that are similar but not identical; keep the prompt and change only the seed, five times. The post on seeds explains the same idea for images.

Which file format should I choose?

WAV is the default and the right choice for editing in a video or audio tool, since it is uncompressed. FLAC is lossless too and smaller. MP3 and AAC are for the web or a quick preview, and for those two you can set a bitrate between 32 and 320 kbps; left empty, the encoding is automatic.

What can it not do?

It is built for sound effects and ambience, and that is all its description promises: for a song or a line of dialogue, pick a music or speech model instead. It also does not watch your video. Matching a sound to the exact frame where a glass hits the floor is a job for the video to sound version of the same model, Mirelo SFX 2.0 video to audio, which listens to the picture. For another text to sound model with looping ambience, see the sound effects V2 guide, and every audio model is on the models page.