Practical Guides
How to Prepare Your Music for AI Model Training
A personalized model can only learn from what you give it. Most disappointing results trace back to preparation rather than the model: mixed eras of one catalog, files with the wrong rights status, or a folder of rough bounces next to finished masters. This guide is the checklist we would want before pressing train.
Who this is for: For artists and producers about to train a personalized model on their own catalog, whether in ANYANO or another tool.
By ANYANO Editorial · Published · 9 min read
Start with rights, not files
Before thinking about audio quality, sort your material by who controls it. A track you wrote, performed and produced alone is simple. A co-write, a record with a session singer, or a mix built on a licensed sample is not — other people have a stake in it.
ANYANO asks you to confirm you own each file or have permission to use it before upload, and asks for a separate confirmation from the performer before a vocal model is trained. That confirmation is a real decision, not a checkbox to click past. If you are unsure about a track, leave it out: a model trained on fewer files you fully control is more useful than one you can't safely release music from.
- Solo-written, solo-produced tracks: usually fine to include.
- Co-writes and collaborations: ask your collaborators first.
- Vocals by another singer: you need that singer's explicit consent.
- Tracks with licensed or uncleared samples: leave them out, or use stems without the sample.
- Label-owned masters: check your contract. Owning the song is not the same as owning the recording.
Decide what sound you want the model to learn
Personalized models learn tendencies: the harmonic habits, rhythmic feel, arrangement density and tonal balance that repeat across your files. If your catalog spans ten years and four genres, the model tries to average them all, and the result often sounds like none of them.
Pick one direction per model. "My last two EPs" is a better training set than "everything I've ever made". If you want both your ambient and your club work, train two Sounds rather than one.
File quality checklist
Use the best version of each file you have. The model learns what it hears, including artifacts. Low-bitrate MP3 smearing, clipping and dither noise become part of "your sound" if they appear across many files.
- Prefer lossless WAV or FLAC masters over MP3 exports.
- Avoid files that clip or were pushed hard into a limiter just to be loud. A clipping detector or LUFS meter will show this quickly.
- Remove long silences, count-ins and trailing noise.
- Keep sample rates consistent across the set where you can.
- Don't include the same song twice, for example a radio edit and the album version. Duplicates over-weight that song.
Organize the dataset before uploading
A simple folder structure saves time later and makes it obvious what the model learned from. This layout works for most artists:
| Folder | What goes in it | Why |
|---|---|---|
| masters/ | Final stereo mixes of finished songs | Teaches the overall balance and arrangement |
| stems/<song>/ | Drums, bass, vocals and other parts per song | Lets you train on parts, or exclude a part |
| instrumentals/ | Versions without lead vocal | Useful if you don't want vocals learned |
| midi/ | MIDI files for melodies and chords | Keeps note information alongside audio |
| excluded/ | Anything you're unsure about | Keeps doubtful files out of training |
Name and label consistently
Descriptive filenames make a catalog auditable: song title, version and part, for example "night-drive_master_v3.wav" or "night-drive_stem_bass.wav". Tempo and key are worth recording as well. The free BPM & Key Finder can batch-analyze a folder and export a CSV you can keep with the dataset.
Labels also help you later, when a generation sounds off and you want to know which part of the catalog it leaned on.
Final check before training
- Every file has a clear rights status, and doubtful files are in excluded/.
- The set represents one musical direction.
- Files are lossless where possible, not clipped, and have no duplicates.
- Vocals are either included with the performer's consent or removed.
- You've listened through the set once, start to finish, as a listener would.
Key takeaways
- Rights decide what goes in. Audio quality comes second.
- One coherent direction per model beats a whole-career average.
- The model learns artifacts too, so use your cleanest versions.
- A labeled, organized dataset makes results explainable later.
Limitations
- Good preparation improves the odds of a useful model. It does not guarantee a particular result.
- Rights guidance here is practical, not legal advice. Contracts and local law decide who controls a recording.