A DEEP LEARNING BASED PITCH ANALYSIS MODEL TO IDENTIFY PITCH ANOMALIES AND ASSESSING THE HARMONY OF INSTRUMENT AND VOCAL COMPONENTS

Abstract

Due to its intricate pitch structures and gamakas or ornamentations and raga specific rules, Carnatic music face challenges in automated analysis of vocal performances. Researchers have proposed innovative approaches in music information retrieval on categorizing different types of music, identifying instruments and pitch extraction, little research have been made on assessing the harmony of instruments and vocals and also in automated analysis of pitch anomalies of vocal performances. An innovative pitch analysis PA deep learning model is proposed in this paper to detect pitch anomalies in vocal performances and also evaluating the harmony between the vocals and accompanying instrument violin. The proposed PA model incorporates the CREPE deep learning model for fundamental frequency (Sa) extraction directly from time-domain audio signals. The pitch values obtained are mapped to remaining fifteen  notes in Carnatic music and hence pitch anomalies are identified for those notes which are not part of raga’s arohana and avarohana.  From the experimental results, it is observed that the proposed PA model shows good capability in accurately identifying pitch anomalies across multiple Carnatic ragas and assessing accompaniment coherence with the vocals. This research provides a comprehensive tool for music analysis, with potential applications in music education, performance evaluation, and intelligent music tutoring systems.