Acoustic Modeling for Automatic Lyrics-to-Audio Alignment

• DOI: 10.21437/interspeech.2019-1520
• Publication Date: 2019-09-13T20:32:51Z
• Authors: Chitralekha Gupta; Emre Yılmaz; Haizhou Li
• Keywords: Lyrics

Automatic lyrics to polyphonic audio alignment is a challenging task not only because the vocals are corrupted by background music, but also there lack of annotated corpus for effective acoustic modeling. In this work, we propose (1) using additional speech and music-informed features (2) adapting models trained on large amount solo singing towards music small in-domain data. Incorporating information such as voicing auditory together with conventional aims bring robustness against increased spectro-temporal variations in vocals. By model data, reduce domain mismatch between training testing We perform several experiments present an in-depth error analysis features, adaptation techniques. The results demonstrate that proposed strategy provides significant reduction word boundary over comparable existing systems, especially more data long-duration musical interludes.