Machine Learning of Timbre Features for The Western Chinese Free-Reed Instrument Hulusi (de)
* Presenting author
Abstract:
The hulusi, played in China's Yunnan Province, is a free-reed pitched wind instrument, an instrument kind not present in the West. It consists of a mouthpiece, a gourd, and three bamboo tubes, all with free reeds made of copper. The principal bamboo tube located in the middle has seven finger holes. Therefore, the tube length determines the instrument pitch, not the free-reed eigenfrequency, different from e.g. the Western accordion or blues harp. This study utilizes a machine learning model, implemented in Comsar framework (https://github.com/ifsm), to investigate timbre features of the hulusi, focusing on clustering and recognition of instruments and articulation. Measurements of different hulusi pitches of C, B-flat, A, G, and F are taken and analyzed with respect to spectral centroid, spectral spread and spectral flux. A self-organizing Kohonen map (SOM) is trained using these timbre features, resulting in clusters of instruments and estimations on articulatory variability.