AI music enhancer for Suno & Udio

Traditional audio enhancers (EQ + compression + denoise) assume the input is "clean signal plus additive noise". AI-generated music contains structural codec artifacts that violate this assumption — generic enhancers can amplify them. This tool uses a different approach: a model trained on codec-artifact restoration specifically.

Apollo is a neural network published at ICASSP 2025 (Li et al., "Apollo: Band-sequence Modeling for High-Quality Audio Restoration"). It was trained to restore MP3-compressed music — specifically MP3s in the 24-128 kbps range — toward lossless quality. The same architecture transfers well to neural-codec artifacts from SUNO free-tier output and other low-bitrate generators.

Apollo is trained on music. For speech cleanup (interviews, podcasts), use our Audio Enhancer or Voice Enhancer tools — they target speech-specific noise patterns.

Its length, almost entirely — the model works through the song from start to finish and does the same amount of work per second of audio. A three-minute song restores in about half a minute on a recent phone and faster on a desktop. What does not change the time is how damaged the file is: a badly compressed export and a nearly clean one cost the same, because the network looks at every part either way.

Heavily-compressed music in Apollo's training range — that means SUNO free-tier MP3 (nominally 128 kbps, sometimes measured higher as VBR), old web MP3s at 96–128 kbps, and Riffusion MP3 exports. Near-transparent inputs gain less: Udio free already ships 320 kbps MP3, SUNO Pro ships 320 kbps or WAV, and the authors note results above 128 kbps are outside the model's evaluation range. Note that SUNO output also carries a ~16 kHz spectral ceiling baked in by its training data, which restoration alone cannot rebuild. Lossless inputs (FLAC, WAV from a DAW) gain little; heavily damaged material (clipping, dropouts) is also out-of-distribution — try the Audio Restoration tool first.

No. The restoration runs in your browser — your audio file is not sent to us. We log only anonymous performance events (model load time, inference duration) to spot bugs and improve speed; see the Privacy Policy.

Yes. The artifact profile from neural audio codecs is similar across these tools — Apollo handles all of them. Best results on the same family of damage it was trained on (MP3-style codec loss).

Not MP3 again, if you can avoid it. The whole point of the run was to undo what MP3 compression did, and saving the result back to a lossy format immediately spends part of what you recovered. WAV or FLAC keeps it whole for further work or for archiving; a lossy export makes sense only at the very end of the chain, when the file is going straight to a listener and nothing else will touch it.