Application of VGGish and YAMnet Model for North Indian Raga Music Recognition using Transfer Learning

Raga is the foundational framework for Indian classical music (ICM), consisting of successive notes. This same principle applies to both the Carnatic and Hindustani Classical music traditions. The significant benefits of recognizing ragas from audio go beyond Music Information Retrieval (MIR), which includes content filtering, teaching/learning, and music therapy. This research paper demonstrates a method for recognizing Hindustani classical ragas using transfer learning with pre-trained models VGGish and YAMnet from raw audio spectrograms. VGGish was particularly developed to extract audio features or embeddings from raw audio streams, whereas YAMNet is a deep-learning model designed for identifying and classifying a wide range of sound events. These models are useful for audio analysis and understanding across a wide range of applications. The proposed techniques obtained an overall testing accuracy of 96.88% for VGGish and 93.3% for YAMnet on a dataset of 200+ samples of 12 ragas played on four distinct musical instruments. The standard confusion matrix is used to determine the study results for these models.