Understandable and Singable Musical Lyrics Translation

Anonymous

Abstract: Translating lyrics in musicals is a new and challenging task due to various constraints to consider. While previous song translation works explore ways to incorporate and satisfy music constraints, they cannot ensure basic translation quality which is critical in musicals. This paper is dedicated to enhancing translation quality while simultaneously maintaining the singability features, such as adherence to length and rhyme constraints. Our approach consists of three main components. First, we collect a dataset to train reward models, giving an automatic evaluation of translation quality. To enhance both the singability and translation capabilities, we adopt a two-stage training recipe with filtering techniques. Finally, our inference-time optimizing framework composes the whole-song translation. Extensive experiments of both automatic evaluations and human evaluation not only show improvements over the baseline method by large margins but also demonstrate the effectiveness of different components in our approach.

Paper Type: long

Research Area: NLP Applications

Contribution Types: NLP engineering experiment

Languages Studied: English, Chinese