Word-Level Lyrics Alignment Annotations for the MUSDB18 Test Set

This dataset provides word-level timestamp annotations for the MUSDB18 test set, consisting of 45 professionally produced English songs spanning a range of musical genres. Originally developed as a benchmark for music source separation, MUSDB18 also includes metadata such as pitch and vocal activity. The addition of precise word-level lyrics alignment enables its use for evaluating automatic lyrics transcription (ALT) systems. Annotations were created by a professional music producer using source-separated vocals and preliminary alignments generated by an existing lyrics alignment model. Each word’s timestamp was manually adjusted and the lyrics were edited where necessary to ensure accuracy. The annotation process took approximately 80 hours. This dataset serves as an alignment extension of the MUSDB18 lyrics extension and is intended to support research in lyrics alignment, singing voice analysis, and multimodal music processing.