IEEE ICONET26
MelodicVision (Melvis)
AI music genre classifier, presented at an IEEE conference
PyTorchResNet-18librosaGrad-CAMTransfer learning
The problem
GTZAN is tiny (799 training clips). Training an audio model from scratch overfits; the question was whether a vision backbone could transfer what it knows about edges and textures to sound.
What I built
- Audio → Mel spectrogram pipeline with librosa; ResNet-18 fine-tuned with discriminative layer-wise learning rates so early filters keep their ImageNet knowledge.
- Grad-CAM explainability: the model attends to harmonically rich mid bands for classical and jazz, broadband energy for metal.
- Sliding-window ensemble so full-length songs of any duration get classified, not just 30-second clips.
Result
82% test accuracy with only 799 training clips. Accepted and presented at ICONET26 (IEEE Computer Society Lebanon Chapter), co-authored with Dr. Ziad Doughan.
Hear what the model sees
MelodicVision turns sound into a Mel spectrogram and reads it like an image. Use your microphone, or play a synthesized tune, and watch the spectrogram draw itself live.
highmidlow
Mel · 64 bands · live
Runs in your browser only; nothing is recorded or uploaded.
Gallery
