ChordSeer

The objective was to build a neural network capable of identifying 8 distinct guitar chords in realtime. My journey began with a standard MLP model using averaged MFCC features, but I quickly hit a performance ceiling at ~51% accuracy. The model suffered from severe overfitting, proving that simple feature averaging destroyed critical temporal data; Dropout layers didn't help, they just stopped the learning process entirely. The breakthrough came when I shifted my strategy to treat audio as visual data. I rebuilt the pipeline to generate Mel Spectrograms (converted to dB) and fed them into a custom 2D CNN architecture. By optimizing preprocessing parameters specifically setting the sample rate to 16kHz and implementing LayerNormalization with L2 regularization, I solved the overfitting issue. The final model achieved a massive leap in performance, reaching 98.75% test accuracy.