Detection of musically relevant regions in multiresolution time-frequency representations evaluated on piano recordings

This paper investigates possible estimators of musical information in subregions of a time-frequency representation of an audio signal, in the context of obtaining multiresolution time-frequency representations through iterative refinements. An experiment was conducted comparing estimators extracted from STFT spectrograms of Disklavier performances to reference values extracted from the corresponding MIDI files. Different reference values were considered, capturing the presence of musical information in the time-frequency plane that are relevant to music information retrieval tasks such as automatic music transcription and timbre analysis. The impact of the size of the timefrequency subregions and initial resolution of the STFT were analyzed. The influence of the introduction of simple energy decay models in the reference values was also investigated. A second experiment was conducted evaluating chosen estimators as features in a predictive model for the binary detection of musically relevant regions of a time-frequency representation. Naive-Bayes models were trained using binary piano rolls (with and without harmonics) as ground truth. Results show that it is possible to detect musically relevant regions of a time-frequency representation with satisfactory results using Shannon and Rényi entropies.