Semitone shift detection in music audio

We present two datasets and an initial algorithm for the classification of semitone shifts. The semitone shift is a pattern in harmonic progressions, which has appeared relatively recently in Western tonal music. It is typically characterised by a change of key by one or two, sometimes more, semitones upwards and the repetition of a section of the music before and after the shift, which typically takes place towards the end of the piece. It is of musicological interest to recognise these key shifts automatically from audio recordings. The ultimate goal is to analyse large data collection to understand historically when this phenomenon started and how it spread and developed. The two datasets comprise popular music with annotations on whether and where the audio recordings contain a semitone shift. The initial algorithm presented is based on self similarity and dynamic time warping (DTW) for the detection of semitone shifts, and it shows encouraging performance on the presented datasets.