Using Deep Learning to Annotate Karaoke Songs

Karaoke is a game in which players sing over pre-recorded instrumental backing tracks. To help the singer, lyrics are usually displayed on a video screen. The synchronization between the lyrics display and the song record, often done manually, is a tedious and time-consuming task. Automation of the annotation of karaoke songs can help save time and effort. In this thesis we use the representation of songs as spectrograms to detect singing times. This timing information can be used later to align the lyrics display with a sound track. Convolutional neural networks are trained to detect at any moment in a song whether the artist is singing or not.