The absence of an automatic page turner that can localise a musician performing to the appropriate section of a music score poses problems. Particularly for musicians with visual impairments who must zoom into the sheet music to distinguish notes, greatly increasing the frequency of page turns. The paper proposes a novel adaptation of the Shazam algorithm and its variations such as PANAKO. These algorithms are used to map a subset of a song to the entire reference song, instead of finding a song in a database of songs. These methods are modified to account for pitch and speed dilations.
In order to disambiguate repetitions and take consecutive recordings into considerations, we use a method used to localize a mobile robot in a known map. In our case the position of the robot in the map would correspond to the time in a song we are in. The strategy chosen is the MonteCarlo localization method, which estimates the posterior probability of the position given the observations by maintaining and updating a set of particles (possible times).
We test our strategy in our self-built test set. Labelling data was time consuming.
Please read the report for a more in depth explanation