Comparison of machine and human recognition of isolated instrument tones

• Published in: The Journal of the Acoustical Society of America 2002-05, Vol.111 (5_Supplement), p.2417-2417
• Main Author: Fujinaga, Ichiro
• Format: Article
• Language: English
• ISSN: 0001-4966; 1520-8524
• DOI: 10.1121/1.4778250

This paper describes three different machine recognition experiments and a recently conducted human experiment in order to compare the abilities of machines and humans to recognize isolated instrument tones. The computer recognition software is based on the Lazy Learning Machine, which is an exemplar-based learning system using a k-nearest neighbor (k-NN) classifier with a genetic algorithm to find the optimal set of weights for the features to improve its performance. The performance of the software was progressively improved by adding more features. These include centroid and other higher order moments, such as skewness and kurtosis, the velocities of moments, spectral irregularity, tristimulus, and time–domain envelope shape. Also, realtime recognition is now possible by using Miller Pucketts PD, a realtime software synthesis system, and his fiddle∼ object. The training data was taken from the McGill Master Samples. The human experiment involved eighty-eight conservatory students. Although the average human scores are similar to the machine scores, the best human subjects far exceeded the capabilities of the machine. The excellent performance of the humans in this experiment presents new challenges for timbre-recognition computer models.