Abstract
In this work, deep learning is utilized to develop a scalable and robust convolutional neural network-based music genre predication. Here, the audio data are divided into overlapping chunks. Then it is converted into Mel spectrograms and log-scaled representations which capture visually relevant frequency content. These spectrograms are resized and fed into a deep CNN architecture. The model has been trained for real-time prediction to guarantee practical usability. It has been designed to provide genre classification of unseen audio files, even though those are in different formats like MP3 and WAV. It is achieved through dynamic generation of Mel-spectrogram and aggregation of chunk-wise prediction. Overlapping temporal chunking, high-resolution Mel-spectrogram processing, and a deep CNN process built for short audio clip classification boost model generalization and prediction robustness. For performance analysis the GTZAN dataset has been used. Performance comparison highlights that proposed model outperforms baselines considering recall, precision, F1-score and accuracy.