Raga Recognition of Indian Classical Music Based on Audio Processing and Stacking Based Ensemble Learning Algorithm

Raga identification is a significant issue in the field of Indian art music since ragas are essential to the composition and performance of the music and are vital to its preservation, education, and retrieval. There aren’t many studies that have looked into this task using techniques like Machine Learning (ML), signal processing, or, more recently, Deep Learning (DL). All of these researches, however, leave open a crucial question: do these ML/DL techniques learn and comprehend Ragas similarly to human experts? Furthermore, a major obstacle to this research is the lack of a large number of rich, labeled datasets, which is what motivates these ML/DL-based techniques. Advanced techniques like deep learning-based Bahdanau Attention-augmented Bidirectional LSTM (BAA-BiLSTM) is proposed to mitigate these drawbacks by better capturing raga nuances. Initially, audio recordings of various ragas from a music collection are collected and pre-processed using Particle Filter (PF) to minimize noise and Double Side Band Amplitude Modulation (DSBAM) for maintaining equal frequency and amplitude. The audio is segmented using Non-Stationarity-Based Adaptive Segmentation (NSAS) to separate pitch and vocals from noise and silence. These segmented signals features are extracted using Gammatone Frequency Cepstral Coefficients, Code Excited Linear Prediction, spectrum flux, short-term energy, and Recurrence Quantification Measure. Then the features are further given to the deep learning ensemble classifier which contains Bahdanau Attention-augmented Bidirectional LSTM (BAA-BiLSTM) meta-classifier with Recursive Tensor Neural Network (RTNN) and Jordan Neural Network (JNN) to identify ragas accurately. As a result of proposed model accuracy is 98.4%, precision is 92.7% respectively. Consequently, this method is highly suitable for real-time applications to identify Indian classical music’s ragas based on audio signals.