Deep Learning for DJs

Audio features are first extracted using Librosa and NumPy: Mel-Frequency Cepstral Coefficients, Zero Cross Rates, Root Mean Squares, Spectral Rolloffs, and Chromagram values across each 10-second chunk of an audio sample, along with their means and variances.

TinyTag extracts metadata from downloaded files, which is used to pull lyrics for each song through the LyricsGenius API. Lyrics are cleaned (stopwords and irrelevant punctuation removed) and vectorized using the OpenAI Large Embeddings model into a final stored data frame.

Librosa code extracting MFCCs, zero-crossing rate, RMS, spectral rolloff, and chromagram per audio segment.

TinyTag metadata extraction and LyricsGenius API lookup code.

Regex cleaning and tokenization of raw lyric text.

Batched OpenAI embeddings API calls over the cleaned lyric text.

The final stored data frame: 422 tracks with audio features, lyrics, and text embeddings.

Central tendencies of key audio features are calculated to identify differences, with Kernel Density Estimate graphs plotted per feature and separated by playlist to visualize density differences.

Correlation heat maps identify covariances between key audio features to inform feature engineering, and a principal component analysis visualizes distinctions between playlists in terms of audio features.

Summary statistics for the audio feature set across all 422 tracks.

Density distributions of each audio feature, split by playlist.

Correlation heatmap of the audio feature set.

A t-SNE is conducted on 3 principal components of the lyrical features across the Dance, Hindi, and Latin playlists. Distinctions are observed, particularly between Hindi, Dance, and Rap music.

Word cloud of the most common lyric tokens across the dataset.

A second word cloud highlighting multilingual lyric tokens across playlists.

A third word cloud, dominated by Spanish and Portuguese Latin-playlist lyrics.

Playlists are label-encoded, train/test splits are created, and each feature is scaled to prevent unfair influence.

Logistic Regression classifies audio features with better-than-random results; Gradient Boosted Trees are tuned across learning rates to reduce overfitting and increase test accuracy.

Label encoding, train/test split, and feature scaling for the supervised models.

Logistic regression baseline: roughly 56% train/test accuracy.

Gradient boosting classifier swept across learning rates to find the best validation accuracy.

Model comparison table: adding textual features lifts validation accuracy from 64% to 85%.

A 7-layer neural network is built in PyTorch, trained with cross-entropy loss, ReLU activation, and a 0.3 dropout rate, using an Adam optimizer and class-weighted batches for fair evaluation.

Early stopping prevents overfitting; results are evaluated playlist-wise and via a confusion matrix.

PyTorch tensors, balanced class weights, and the training DataLoader for the final neural network.

The final neural network architecture: a 6-layer MLP with batch norm and dropout.

Weighted cross-entropy loss, Adam optimizer, and a plateau learning-rate scheduler.

The core PyTorch training loop: forward pass, loss, backward pass, optimizer step.

Early-stopping logic that checkpoints the best model by test accuracy.

Final per-playlist classification report: 84.71% test accuracy.

Confusion matrix for the final playlist classifier.

Gaussian Mixture Models explore whether the corpus clusters differently than its supervised categorization. Clusters prove less distinct than expected, with songs from different playlists grouped together.

Given the model's high F1-score, it successfully classifies songs into genres based on specific DJ preferences, saving substantial time spent analyzing and categorizing songs. The implementation was limited by available data and compute.

Open questions: which additional features cause unsupervised clustering to diverge from supervised categories? Can the algorithm decide which songs to download in the first place? Can it recommend which tracks to mix, not just classify them?