A theory of musical expectations based upon a set of computational models and human subject experiments.
Studies of brain activation of music cognition with emphasis on structural segmentation, language and music, and expectations.
Pairs of musicians were placed apart in isolated rooms and asked to clap a rhythm together. Each person monitored the other's sound via headphones and microphone pickup was as close as possible. Time delay from source to listener was manipulated across trials. Trials were recorded and clap onset times were measured with an event detection algorithm. Longer delays produced increasingly severe tempo deceleration and shorter delays (< 11.5 ms) produced a modest, but surprising acceleration. The study's goal is to characterize effects of delay on rhythmic accuracy and identify the region most conducive to ensemble playing. The results have implication for networked musical performance. Network delay is a function of transmission distance and / or internetworking (routing) delays. The findings suggest that sensitive ensemble performance can be supported over rather long paths (e.g., San Francisco to Denver at about 20 ms, one-way). The finding that moderate amounts of delay are beneficial to tempo stability seems, at first glance, counterintuitive. We discuss the observed effect.
Reference:
We investigate a neural network model to explain nonlinear behavior of our auditory system, particularly of basilar membrane. When the input stimulus is composed of two sinusoidal primary tones, we not only hear those original tones, but also perceive many other distortion products, called combination tones, due to nonlinearity inherent in the cochlear mechanism. Among these combination tones, the cubic difference tone at frequency is of particular interest because it is above masking threshold, and thus certainly audible, especially when the intensities of the primary tones are high. However, its amplitude behavior is so unusual that it does not follow the classical square-law at all, and therefore it is very difficult to estimate the amplitude of the CDT. Furthermore, it is also dependent upon the amplitudes and frequencies and the frequency separation of two primary tones. In [1], we have used a 3rd-order Volterra kernels to estimate the amplitude of the CDT, and compared the results with measured data seen in [2]. Even though the results show very close match with some experimental data, it fails to work with other data sets. We therefore consider a neural network model, and investigate the possible feature sets that work best and most efficiently with a neural net.
References:
© Copyright 2005 CCRMA, Stanford University. All rights reserved. |