Article
Formatting-aware Automatic Lyrics Transcription: a Case Study on German Songs (en)
* Presenting author
Abstract:
In this work, we look at state-of-the-art methods for automatic lyric transcription using machine learning models. We provide a brief overview of the current systems, highlighting their strengths and weaknesses, and discuss the significance of separating the vocals as a preliminary step.Many of these systems are primarily tuned to achieve low word error rates, and we believe that they fail to consider the intricate details of written song lyrics. This could lead to a misalignment between the transcribed lyrics on one side and how the music is conceived and experienced on the other. For example, the absence of line breaks makes it harder to perceive the original rhythm, emotional emphasis, or rhyme scheme.Specifically German songs pose distinct difficulties due to the complex grammar, frequent compound words, and strong dialects, thus making it a challenging linguistic landscape. We present results from our system, which not only accurately transcribes lyrics but also addresses other aspects such as punctuation, line breaks, spelling, background singing.In order to evaluate our system, we present a complete revision of a reference dataset specifically designed for the evaluation of multilingual song text transcriptions, as well as an open-source implementation of the lyrics transcription metrics.