Hi everyone! I recently finished a project that allowed me to thoroughly explore the intersection between signal processing and **deep learning.
The goal was clear: to create an AI capable of classifying music genres, not by analyzing the audio directly, but by “teaching it to see” the music through spectrograms.
️ How did I do it? The workflow:
- Data Engineering: I used the GTZAN dataset and developed my own script based on yt-dlp to expand the corpus with contemporary genres such as Trap, Lofi, and Reggaeton.
Critical Preprocessing: I implemented a pipeline with Librosa to normalize the audio, trim silences, and transform the waveforms into 128/128px Mel spectrograms. The music was officially converted into RGB images!
CNN Architecture: I designed a Convolutional Neural Network (CNN) with Conv2D layers, MaxPooling to reduce dimensionality, and Dropout to prevent overfitting.
Results: I achieved an overall accuracy of 80%. The most satisfying part was seeing how the model reached an F1-Score of 1.00 for the Lofi genre, validating the quality of the custom dataset.
Lessons learned:
It’s not always about having the most complex model. I learned that data cleaning and standardization (such as precisely trimming clips to 30 seconds) are just as critical as the network architecture.
Although I initially planned to do fine-tuning, I decided to first consolidate the base architecture to ensure solid learning of the fundamental features of the new genres. We’re still learning!