neerajs7/AST-audio-classifier

🎵 AST Music Genre Classifier

Audio Spectrogram Transformer model fine-tuned for music genre classification on noisy mashup data.

Model Description

This model is a fine-tuned version of MIT's Audio Spectrogram Transformer (AST) with custom Patchout regularization for robust music genre classification.

Key Features

• 🎯 Classifies 10 music genres with high accuracy
• 🔊 Robust to noise, tempo variations, and audio mashups
• 🚀 Achieves 98.19% F1-Score on public leaderboard
• 🎨 Custom Patchout regularization for better generalization

Performance

Metric	Score
Public Leaderboard	0.98191 Macro F1
Private Leaderboard	0.97634 Macro F1
Competition	Kaggle Messy Mashup Genre Classification

Model Architecture

• Base Model: MIT/ast-finetuned-audioset-10-10-0.4593
• Custom Components:
  • Patchout regularization (time=0.4, freq=0.2)
  • Custom classification head with LayerNorm and GELU activation
  • Dropout layers for regularization
• Parameters: ~86.5M trainable parameters

Architecture Details

Input (Audio) → AST Encoder → Patchout → Pooling → Classification Head → Genre (10 classes)

Classification Head:

LayerNorm(768) → Dropout(0.15) → Linear(768→384) → GELU → Dropout(0.1) → Linear(384→10)

Citation

@misc{ast-genre-classifier-2026, author = {Neeraj Surin}, title = {AST Music Genre Classifier with Patchout Regularization}, year = {2026}, publisher = {HuggingFace}, journal = {HuggingFace Model Hub}, howpublished = {\url{https://huggingface.co/neerajs7/AST-audio-classifier}} }

Acknowledgments

• Base AST model: MIT/ast-finetuned-audioset-10-10-0.4593
• Competition: Kaggle Messy Mashup Genre Classification Challenge
• Framework: Hugging Face Transformers

License

MIT License - See LICENSE file for details

Model Card Authors: Neeraj Surin Model Card Contact: https://huggingface.co/neerajs7