AST Audio Classification Model (Messy Mashup)

Introduction

Model Description

• Developed by: Rudransh Mathur
• Institution: Indian Institute of Technology, Madras
• Model type: Transformer-based Audio Classification Model
• Base model: AST (fine-tuned on AudioSet)
• Framework: Transformers (Hugging Face) + PyTorch
• License: MIT

Model Sources

• Repository: https://github.com/rudransmathur/dl-genai-project-26-t1
• Kaggle Competetion: https://www.kaggle.com/competitions/jan-2026-dl-gen-ai-project

Intended Use

• Audio classification tasks
• Music/audio tagging
• Experimental research in audio transformers

Training Details

• Dataset: Messy Mashup Audio Dataset
  • GENRES: [blues, classical, country, disco, hiphop, jazz, metal, pop, reggae, rock]
  • STEMS = [drums, vocals, bass, other]
• Epochs: 10
• Optimizer: AdamW
• Loss Function: Cross-Entropy Loss
• Scheduler: Cosine Sheduler with warmup steps

Preprocessing

• Randomly sampled audio stem files within the same genre and mixed to create a mixed song audio song similar to test dataset
• Added 5 seconds of noise from the noise dataset 2-3 times on a random basis in the audio file.
• Audio inputs converted using AST feature extractor
• Sampling rate aligned with model requirements

Performance

• Best Validation Accuracy: 0.87
• Best Validation Loss: 0.40373
• Best Test Accuracy: 0.92
• Best Validation Loss: 0.3458

🚀 Usage