Refining audio-to-score alignment for singing voice transcription

Note-level automatic singing transcription, which involves extracting both time boundaries and accurate pitch from a singing voice, remains a significant challenge in Music Information Retrieval. Despite advancements in deep learning, progress is constrained by the labor-intensive task of annotating datasets, leading to ongoing data scarcity. To tackle this, we introduce a novel audio-to-score alignment algorithm that effectively synchronizes timed events between score and audio. Our method not only delivers precise alignments but also includes a mechanism to evaluate their reliability. Using this approach, we developed a Singing Onset Labels Extracted Automatically dataset, which facilitated training a model on a diverse range of musical genres and achieved state-of-the-art performance in singing onset estimation for pop music. Both the alignment algorithm and the Singing Onset Labels Extracted Automatically dataset publicly available for use by other researchers.