LyriX : Automated Lyrics Transcription Using Deep Learning-Based Residual UNet Model

Automated Lyrics Transcription has emerged as a crucial research area to enhance music analysis and comprehension. In this paper, we present a novel approach for Automated Lyrics Transcription for English pop and rock songs using a deep learning-based residual UNet model. A comprehensive dataset of annotated songs, comprising lyrics and their corresponding audio tracks, was curated to train and evaluate the model. The performance of the vocal separation model was evaluated using the Signal-to-Distortion Ratio, demonstrating significant improvements over existing methods. Additionally, we propose an innovative lyrics transcriber, leveraging the vocal separation model, and evaluate its accuracy using the Word Error Rate. The results indicate high precision in the lyrics transcription process. Our research contributes to the advancement of automated lyrics transcription techniques, offering valuable insights for further developments in music analysis and application areas.