Abstract
Sheet music has long been regarded as one of the most effective medias for musicians, music players, and amateurs to communicate with each other. It is also an intuitive way for non-professionals to learn how to play a musical instrument or sing a song. However, not all composers have willingness to share their own sheet music, especially those protected by strict copyright regulations. For amateurs or novice musicians without the capability to distinguish the chords by their own ears, it would be difficult for them to enjoy the happiness of playing other's music, or singing songs. In order to provide the beginners with an effective way to play music or sing songs without a given sheet music, we developed a sheet music generator-'DeepSheet', which was engineered based on deep learning techniques. Basically, the DeepSheet system is mainly composed of three distinctive components as follows: 1. voice separation of the audio song, 2. chord estimation of the background music, and 3. alignment of lyrics and music beats. The experimental results drawn from more than 150 songs of the Beatles and Queen indicated that the newly developed DeepSheet system has the capability and sensibility to generate sheet music with accuracy of approximately 76%.