Approaching polyphonic transcription of piano sounds

Abstract

Polyphonic transcription of piano is a challenging task due to the specific characteristics of its sound spectrum. Pattern matching method, which compares the spectrum with a set of spectral patterns, has proven to give good results, although some limitations still exist mainly when analyzing notes of the lower octaves. Present research is oriented to improve both the accuracy of the spectral patterns, especially for lower octaves, and the matching-score calculation. The former is being carried out by developing a model of the vibration of piano. The last is aimed to decide whether a simple internal product or a more complex MSE calculation are needed and how the spectral pattern has to be modeled regarding partial levels. The model of the vibration permits calculating the partial frequencies, for which it takes into account the vibration of the stiff string and the effect of the soundboard impedance. The soundboard impedance modeling is under a deeper study using both real measurements and FEM modeling. As further work, once the model is completed, it is expected to generate the patterns by using a waveguide synthesis method. [Work supported by Spanish National Project TEC2006-13067-C03-01/TCM.]