RajaKanna/cnn14_messy_mashup_deployment

🎵 CNN14 Messy Mashup Genre Classifier

This Hugging Face Space deploys my pretrained CNN14 deep learning model for music genre classification on messy mashup audio clips.

📌 Project Goal

The task is to predict the genre of a mashup audio clip from the following classes:

• —blues
• —classical
• —country
• —disco
• —hiphop
• —jazz
• —metal
• —pop
• —reggae
• —rock

The mashup clips are noisy, mixed, and distribution-shifted audio samples.

🧠 Model Used

• —Architecture: CNN14-inspired audio classifier
• —Input: Raw .wav audio
• —Internal preprocessing:
• —resampling to 32 kHz
• —waveform normalization
• —log-mel spectrogram extraction
• —Inference strategy: 5-crop Test Time Augmentation (TTA)

⚙️ Training Setup

The model was trained on a synthetic mashup dataset created to closely match the test distribution:

• —stems mixed from same-genre songs
• —additive environmental noise
• —random SNR variation
• —random gain perturbation
• —normalization to avoid clipping

This was done to improve robustness under realistic mashup conditions.

🚀 How to Use

1. 1.Upload a .wav mashup audio file
2. 2.The app predicts:
3. 3.Top predicted genre
4. 4.Top-3 genre probabilities

🔗 Model Repository

The deployed model weights are stored in this Hugging Face model repo:

RajaKanna/cnn14_messy_mashup_genre_prediction

🔎 Usage Note

This Space is designed for interactive single-file inference. Upload one .wav mashup audio file at a time to obtain predictions.

For full competition evaluation, batch inference is performed separately in the Kaggle pipeline.

📚 Notes

This deployment is intended for demo, usability, and model serving evaluation as part of an academic deep learning project and its evaluation.