Tonal Similarity from Audio Using a Template Based Attractor Model

A model that calculates similarity of tonal evolution among pieces in an audio database is presented. The model employs a template based key finding algorithm. This algorithm is used in a sliding window fashion to obtain a sequence of tonal center estimates that delineate the trajectory of tonal evolution in tonal space. A chroma based representation is used to capture tonality information. Templates are formed from instrument sounds weighted according to pitch distribution profiles. For each window in the input audio, the chroma based representation is interpreted with respect to the precalculated templates that serve as attractor points in tonal space. This leads to a discretization in both time and tonal space making the output representation compact. Local and global variations in tempo are accounted for using dynamic time warping that employs a special type of music theoretical distance measure. Evaluation is given in two stages. The first is evaluation of the key finding model to assess its performance in key finding for raw audio input. The second is based on cross validation testing for pieces that have multiple performances in the database to determine the success of recall by distance.