📄 Arabic Music Classification and Generation using Deep Learning

Summary

The authors study two tasks on Egyptian instrumental music: identifying the composer of a recording and synthesizing new audio in a similar style. They assembled their own collection of 256 instrumental tracks spanning classical Egyptian composers (Mohamed Abdelwahab, Reyad ElSonbaty, Balegh Hamdy) and modern ones (Waleed Saad, Mohamed Rahem, among others). For classification, tracks are converted to WAV, volume-normalized, sliced into short overlapping windows, and turned into fixed-size spectral features (MFCC, later log mel spectrograms) that feed a CNN. Five successive experiments progressively rebalance the composer set, remove a confusing composer, switch features, and add data augmentation. The best configuration reaches about 81.4% test accuracy. For generation, an autoencoder (described as a CNN autoencoder and elsewhere as a VAE) encodes mel spectrograms into a latent space. New samples are decoded and converted back to sound. The authors admit the output only loosely resembles the source and contains noise. The work matters mainly as an early, self-collected benchmark for Arabic music.

Key points

• Dataset: 256 Arabic instrumental tracks, 2.5 to 50 minutes long. Experiment 1 used 11 composers with an uneven split; experiment 2 narrowed this to 9 better-represented composers.
• Balancing rules: no more than 20 clips per composer and no more than 145 segments (about 5 minutes) per clip. Hamza Namira and Waleed Saad were dropped, and later Omar Khorshid, whose tracks are mostly reworked versions of others' pieces.
• Windows shrank from 10 s to 5 s, both with a 2 s hop. MFCC inputs were zero-padded to (20, 938) and then to (20, 469). The data split was 70:15:15 for train, test and validation.
• Augmentation with pitch shifting and time stretching gave 0.7316 test and 0.6935 validation accuracy. Adding frequency masking raised this to 0.8143 test and 0.8458 validation, with 0.9697 training accuracy.
• Generation loss went down during training, but the decoded audio was noisy. The authors conclude the autoencoder failed to capture the complexity of the compositions.

Interesting ideas

• I1. Remove a class whose items are largely derivatives of other classes (Khorshid's covers), since it injects systematic label noise rather than a distinct signal. — For Mova (learn): When building silver treebanks, drop source genres or annotators whose output mostly reworks another source's conventions, so witness agreement is not inflated by near-duplicates.
• I2. Cap contributions per class and per item (20 clips, 145 segments) so a few long recordings cannot dominate training. — For Mova (learn): Limit sentences taken from any single document or text genre in silver data, so the parser does not overfit the style of one long fable or textbook.
• I3. Add augmentations one at a time and measure each step, which showed that frequency masking gave the largest jump. — For Mova (learn): Evaluate each training-data perturbation (token dropout, synonym swaps, clause reordering) as its own ablation step against a fixed negative control before stacking them.
• I4. Read the confusion matrix for pairs of stylistically close classes (ElSonbaty vs. Hamdy) to judge whether a change really sharpened discrimination. — For Mova (tree): Track confusion between neighbouring dependency labels (obj vs. iobj, nmod vs. obl) per experiment rather than only LAS, to see whether a fix targets the hard distinctions.
• I5. Masking whole bands of the input forces the model not to lean on a narrow feature subset. — For Mova (tokens): Randomly hide suffix or capitalization features during tagger training so morphology decisions stay robust when one cue is missing or misleading.