Chroma-based estimation of musical key from audio-signal analysis

This paper deals with the automatic estimation of key (keynote and mode) of a music track from the analysis of its audio signal. Such a system usually relies on a succession of processes, each one making hypotheses about either the signal content or the music content: spectral representation, mapping to chroma, decision about the global key of the music piece. We review here the underlying hypotheses, compare them and propose improvements over current state of the art. In particular, we propose the use of a Harmonic Peak Subtraction algorithm as a front-end of the system and evaluate the performance of an approach based on hidden Markov models. We then compare our approach with other approaches in an evaluation using a database of 302 baroque, classical and romantic music tracks.