Multiple pitch transcription using DBN-based musicological models

International audienceWe propose a novel approach to solve the problem of estimating pitches of notes present in an audio signal. We have developed a probabilistically rigorous model that takes into account temporal dependencies between musical notes and between the underlying chords, as well as the instantaneous dependencies between chords, notes and the observed note saliences. We investigated its modeling ability by measuring the cross-entropy with symbolic (MIDI) data and then proceed to observe the model's performance in multiple pitch estimation of audio data