Probabilistic Classi cation of Monophonic Instrument Playing Techniques

1. Introduction Understanding the underlying intentions of a music performer is crucial to enable a machine such as an automated accompaniment system to interact intelligently with a musician. Particularly, understanding the symbol associated with a tone the player generates allows a machine to create response that is in concordance with the symbol. We dene playing technique as a symbol such as the expression marking that reects the intention of a human performer in a per-ceivable change in timbre. We believe that it is crucial to understand that some playing techniques are inherently ambiguous, and to associate with an input signal the degree of ambiguity along with the estimated class labels. For example, if the class of dynamics are described by “soft” and “loud,” it is irrelevant to ask whether a “moderate” sound is “soft” or “loud”-it only makes sense to say that it either belongs to both or to neither. We shall express ambiguity by modeling a set of playing techniques as a posterior distribution, and using statistics obtained from the distribution to determine the ambiguity. We shall group playing techniques that acts on a same, continuous quality into one set that in turn generates a posterior distribution as shwon in Figure 1. Particularly, we hypothesize that ambiguous sounds are the main cause of misclas-sication, and such sounds creates a distribution with a high variance. Existing research in detecting the playing technique involved discretization of techniques involving a continuous factor and converting it into a problem of classication. For example, the position of the bow on a violin was discretized by recording two points and choosing one of the two positions [1]. Other research exclusively dealt with playing techniques of discrete quality such as whether a bass guitar string was slapped or not [2]. Both approaches did not associate any posterior probability with the output label, and thus suffered when recognizing notes that even humans have trouble distinguishing [1]. Another approach involved extraction of “perceptual” features that were used to control another musical instrument [3]. This approach did not attempt to symbolize timbre into playing techniques. In this paper, we modeled a set of playing techniques that acts on a same continuous factor (position of the bow on a violin) as a posterior distribution given the input signal using a hybrid of Gaussian Mixtures (GMM) and Relevance Vector Machine (RVM). We then rejected data whose variance exceeds a threshold given …