Lyrics Are Meant to Be Sung: Modeling Singable Similarity for Cover Song Identification with lyrics

ACL ARR 2026 January Submission9346 Authors

Keywords: Singability modeling, Cover song Identification, Lyric Transcription

Abstract: Lyric similarity is commonly modeled through various NLP models, implicitly treating lyrics as ordinary text. However, lyrics are written to be sung, and their similarity is fundamentally constrained by singability: whether different sentences can be performed over the same melody. We reconceptualize lyric similarity as singable similarity and propose a framework that learns lyric representations grounded in this principle. Our approach leverages lyric translation pairs that are optimized to fit the same melody, providing natural supervision for learning sentence-level similarity under singability constraints. We introduce a singability encoder model that jointly captures semantic content and syllable structure. Evaluated on line-level translated lyric retrieval and lyric-based cover song identification (CSI), our singability-aware representations consistently perform well on lyric tasks. These results highlight singability as a crucial dimension for lyric-based music information retrieval.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: NLP Applications

Contribution Types: Model analysis & interpretability

Languages Studied: English, Korean, Japanese,Spanish,German,French

Submission Number: 9346