Upgraded Genre Detector

It's been ~2 years since What's my Genre? was first rolled out, and I've learned a lot while building the AI Song Checker.

I decided I'd take all of that knowledge and attempt to "extend" the existing MAEST Model with a bunch of genres that I felt were missing, including Solo Piano, Indie Folk, Phonk, and Organic House.

The end result wasn't too bad.

What stood out

• Clear, distinctive sounds did best. Solo Piano, Study beats, Cinematic, and Organic House
• Electronic styles with strong signatures held up well. Phonk, Amapiano, DnB
• Overlapping indie and alternative labels sucked. Alternative Dance, Indie Dance, Folktronica
• Distribution: 11 excellent, 21 good, 31 moderate, 22 need work
• Mean AP 0.437 across all 85

Technical approach

• About 500,000 audio samples across all genres (this part took ages)
• Stacked classifier, frozen base, new MLP head
• About 300 epochs on cached logits
• Zero drift on the original 519 by design
• Feature adapter helped alignment

The 85 new genres

Sorted by average precision.

Top performers (AP ≥ 0.70)

• Solo Piano 0.899
• Study beats / Jazz-hop / Chill-hop 0.841
• Cinematic / Epic Music 0.762
• Heavy Drum & Bass 0.761
• Melodic Metal 0.751
• Lo-fi Hip-Hop 0.744
• Light Drum & Bass 0.742
• Phonk 0.736
• Afrobeats / Afro-pop / Afro-fusion 0.726
• Amapiano 0.719
• Solo Guitar 0.714

Strong performers (0.50–0.70)

• Brostep 0.664
• Future Bass 0.656
• Tribal / Afro House 0.657
• Melodic Techno 0.622
• Acid Techno 0.617
• Deutschrap 0.610
• Detroit Techno 0.603
• Skate Punk 0.596
• Organic House 0.589
• Son Cubano 0.580
• Symphonic Metal 0.576
• UK Hip-Hop 0.572
• Minimal House 0.568
• Latin Hip-Hop 0.567
• Latin House 0.561
• Rap 0.557
• Latin Pop 0.546
• Drill 0.541
• Classical Strings 0.530
• Riddim / Tearout 0.511
• Cyberpunk / Darksynth 0.504
• Trap (EDM) 0.500

Moderate performers (0.30–0.50)

• French Hip-Hop 0.495
• Big Room 0.494
• Latin Rock 0.472
• Acoustic Folk 0.480
• Lo-fi House 0.450
• Electro Pop 0.449
• Slap House / Brazilian Bass 0.446
• Future House 0.444
• Rap Metal 0.435
• Indie Folk 0.425
• Electronica 0.409
• Alt Pop 0.402
• Moombahton 0.400
• Leftfield Bass 0.398
• French / Filter House 0.376
• Alternative / Indie R&B 0.375
• Wave 0.374
• 80s Rock 0.367
• Bedroom / Lo-fi Pop 0.367
• Hyperpop 0.356
• Folk Pop 0.349
• Folk Punk 0.342
• Electro Swing 0.342
• Jackin House 0.334
• Old-school Hip-Hop 0.339
• Chill House 0.328
• Melodic Hardcore 0.326
• Melodic Dubstep 0.321
• Melbourne Bounce 0.320
• Happy Rave 0.309

Needs improvement (< 0.30)

• Screamo 0.278
• Electronicore 0.266
• Slowcore 0.266
• Post-grunge 0.260
• Global Bass 0.259
• Psydub / Psychill 0.253
• Indietronica 0.237
• Chillstep 0.228
• Chicago House 0.222
• Future Garage 0.218
• 90s Rock 0.204
• Slacker Rock 0.199
• Electro Funk 0.191
• Indie Electronic 0.183
• Vapor Twitch 0.160
• Twee 0.155
• Modern Progressive Rock 0.148
• Chamber Pop 0.147
• Psychedelic / Freak Folk 0.136
• Indie Dance 0.128
• Folktronica 0.121
• Alternative Dance 0.045

Wrap-up

MAEST now knows 604 genres 🎉

Strong gains on distinctive styles. Tougher overlaps still need love. Next up is targeted data and tuning for the weak spots. But for now, I feel like I need a break.