Machine Learning Final Competition

Task Overview 😊

Develop a model to classify song snippets into one of four categories:

• No voices present
• Male voice present
• Female voice present
• More than one person's voice

Datasets

• Train Set: 11886 song snippets with manual labels
• Test Set: 2447 song snippets

Build-from-Scratch Models Overview

• 4-Layer CNN
• ResNet 34

Improvements

• Data Pre-processing ¹
  • Transform mp3 data into Mel spectrogram
• Data Augmentation
  • Time Mask / Frequency Mask / Reverse
  • Mix-up
  • Up-sampling based on labels' weight
  • Human voice separation
• Model Improvement
  • Change the Input layer of ResNet 34
  • Ensembling
• Choice of Hyper-Parameters / Scheduler

Model	Scheduler	Batch Size	Initial lr	Mixup Alpha	Model Weights
CNN 4-layers	MultiStepLR	16	0.001	0.4	-
ResNet 34	ReduceLROnPlateau	16	0.005	0.4	-
ResNet34 + CNN	-	-	-	-	0.5, 0.5

Performance

Model	Epochs	Train Loss	Validation Loss	Validation Accuracy	Test Accuracy
CNN 4-layers	30	0.3828	0.3531	85.79%	76.91%
ResNet 34	24	0.4228	0.2814	88.52%	78.17%
ResNet34 + CNN	-	-	-	-	79.44%

Requirements

pip install torch
pip install librosa

Footnotes

1. See the PDF for details. ↩