Classifying Classical Composers with a CNN-LSTM Model
A hybrid CNN-LSTM model that names the composer behind a piece of classical music from raw MIDI data — 97%+ accuracy across four composers.
Team project for AAI-511 (Neural Networks and Deep Learning) at the University of San Diego, built with Anitra Hernandez and Lakshmi Deepti Pamula. The question: given a piece of music, can a model learn to tell Bach from Beethoven from Chopin from Mozart — not from a recording, but from the raw MIDI data underneath it.
The dataset (Kaggle's midi_classic_music) was badly imbalanced — about 1,020 Bach files against roughly 250 Mozart, 215 Beethoven, and 140 Chopin. We tokenized each composer's files with a REMI tokenizer, used pitch/velocity/duration augmentation to even out the underrepresented composers, then downsampled everyone to 100 files to keep the classes balanced. From there, each MIDI file was rendered to audio with FluidSynth, converted to a mel-spectrogram with librosa, and sliced into time-step images — turning the music into something a CNN could actually look at.
The model itself combines both architectures: four convolutional layers pull feature maps out of each spectrogram image, which then feed into an LSTM instead of a standard fully-connected head, so the model learns how those features change across time rather than treating each slice in isolation. A linear layer and softmax over the four composers close it out.
Trained on 240 images and tested on 80 held-out ones, the combined model hit accuracy, precision, recall, and F1 all above 97%, missing only two files. To check that the hybrid design was actually earning its keep, we also built standalone CNN and LSTM baselines: the CNN alone reached ~90%, and the LSTM alone started at a rough 38% until real optimization work — packed sequences instead of manual padding, a second bidirectional layer, gradient clipping — brought it up to 90% too. The combined CNN-LSTM still beat both outright, heavier preprocessing and all.