Machine recognition of timbre using steady-state tone of acoustic musical instruments
Recent experiments indicate that steady-state portion of an acoustic musical instrument may be sufficient for timbre recognition. Here a computer-based classifier was used to recognize very short samples of steady-state tones. Gregory Sandell's SHARC data comprised of 39 different timbre (23 orchestral instruments, some with different articulations) played at different pitches (total of 1338 spectrums) were used as the data for the Lazy Learning Machine, which is an exemplar-based learning system using k-nearest neighbor classifier with genetic algorithm to find the optimal set of weights for the features to improve its performance.
The features calculated from the spectral data included centroid and other higher order moments, such as skewness and kurtosis.As expected the centroid alone was the best single feature with the recognition rate of 13%, which is much better than chance (2.5%). The best results were obtained using five features: the fundamental, the integral of the spectrum, the centroid, the standard deviation, and the skewness. What was surprising was the recognition varied greatly between instruments. While the French horn and the muted trumpet were recognized near 100%, other instruments did very poorly: such as the oboe (19%), the viola with martele (14%), and the violin with martele (6%). The average overall was 46%.