Piano Timbre Development Analysis using Machine Learning

All 88 keys of a piano at two stages, right after production (stage 1) and one year after playing (stage 2) are investigated using Music Information Retrieval (MIR) timbre extraction and Machine Learning (ML). In \cite{Plath2019} it was found that listeners clearly distinguished both stages but no clear correlation with acoustics, signal processing tools or verbalizations of perceived differences could be found. Using a Self-Organizing Map (SOM) training single as well as double feature sets it is found that spectral flux is able to perfectly cluster the two pianos. Sound Pressure Level (SPL), roughness, and fractal correlation dimension, as a measure for initial transient chaoticity are furthermore able to order the keys with respect to high and low tones. Combining spectral flux with the three other features in double-feature training sets maintain stage clustering only for SPL and fractal dimension, showing sub-clusters for both stages. These sub-clusters point to a homogenization of SPL for stage 2 with respect to stage 1 and a pronounced ordering and sub-clustering of key regions with respect to initial transient chaoticity.