A NOVEL APPROACH FOR SINGER IDENTIFICATION AND VOCAL RANGE ESTIMATION USING HYBRID DEEP LEARNING MODEL

Automatic singer identification and estimating vocal range  of singer are essential  in music information retrieval, with significant applications in music education, digital archiving, personalized training, vocal health monitoring and assisting composers in selecting suitable singers for specific compositions. While previous research focused on singer classification using timbre and spectral features, computational approaches for vocal range estimation of singer  remain underexplored. Identification of singer and analysis of vocal range will have a great impact on commercial and academic domains.  In this research, we propose a hybrid PA (Pitch Analysis) model that employs CREPE deep learning model for pitch extraction, capturing pitch characteristics directly from vocal tracks after isolating it from instruments in a polyphonic recording. The extracted pitch values are then transformed into scalar embeddings and provided as an input to a SRT (Singer Range Transformer) model, which is used to train a Transformer encoder architecture to identify singers and extract their highest and lowest range of pitch. A customized dataset was curated to ensure diversity and robustness, consisting of twenty distinct Carnatic music compositions by the same singer. Experimental results demonstrated that the proposed model achieved high accuracy in identifying the singer and provides reliable vocal range estimations consistent with musicological expectations. This research introduces a novel computational approach to singer identification and vocal range analysis, offering valuable contributions to music education, training, archiving and vocal-health wellness.