Classification of Musical Timbre Using Bayesian Networks

In this article, we explore the use of Bayesian networks for identifying the timbre of musical instruments. Peak spectral amplitude in ten frequency windows is extracted for each of 20 time windows to be used as features. Over a large data set of 24,000 audio examples covering the full musical range of 24 different common orchestral instruments, four different Bayesian network structures, including naive Bayes, are examined and compared with two support vector machines and a k-nearest neighbor classifier. Classification accuracy is examined by instrument, instrument family, and data set size. Bayesian networks with conditional dependencies in the time and frequency dimensions achieved 98 percent accuracy in the instrument classification task and 97 percent accuracy in the instrument family identification task. These results demonstrate a significant improvement over the previous approaches in the literature on this data set. Additionally, we tested our Bayesian approach on thewidely used Iowamusical instrument data set, with similar results. The identification of musical instruments in audio recordings is a frequently explored, yet unsolved, machine learning problem. Despite a number of experiments in the literature over the years, no single feature-extraction scheme or learning approach has emerged as a definitive solution to this classification problem. The ability of a computer to learn to identify musical instruments is an important problem within the field of music information retrieval, with high commercial value. For instance, companies could automatically index their music libraries based on the musical instruments present in the recording, allowing search and retrieval by specific musical instrument. Timbre identification is also important to the tasks of musical genre categorization, automatic score creation, and track separation. This work investigates classification of single, monophonic musical instruments using several different Bayesian network structures and a featureextraction scheme based on a psychoacoustic definition of timbre. The results of this seminal use of graphical models in the task of musical instrument classification are compared with the baseline algorithms of support vector machines and a k-nearest neighbor classifier. Computer Music Journal, 37:4, pp. 70–86, Winter 2014 doi:10.1162/COMJ a 00210 c © 2014 Massachusetts Institute of Technology. Timbre When a musical instrument plays a note, we perceive a musical pitch, the instrument playing that note, and other aspects, like loudness. Timbre, or tone color, is the psychoacoustic property of sound that allows the human brain to readily distinguish between two instances of the same note, each played on a different instruments. The primary musical pitch we perceive is usually the first harmonic partial, known as the fundamental frequency. Pitched instruments are those whose partials are approximate integer multiples of the fundamental frequency. With the exception of unpitched percussion, orchestral instruments are pitched. The perception of timbre depends on the presence of harmonics (i.e., spectrum), as well as the fine timing (envelope) of each harmonic constituent (partial) of the musical signal (Donnelly and Limb 2009).