In investigations for frequency selectivity of the auditory system, the power spectrum model of masking [1] is widely accepted to explain the phenomenon of masking. In this model, it is assumed that when a listener tries to detect a sinusoidal signal with a particular center frequency amid background noise, he makes use of the output of a single auditory filter having a center frequency close to the signal frequency and having the highest signal-to-masker ratio. In addition, it is assumed that the stimuli are represented by long-term power spectra, and that the masking threshold for the sinusoidal signal is determined by the amount of noise passing through the auditory filter. With these assumptions, the power spectrum model can explain masking phenomena such as simultaneous masking. However, this model cannot explain all masking phenomena because the relative phases of the components and the short-term fluctuations in the masker are ignored.
In 1984, Hall et al. have demonstrated that across-filter comparisons could enhance the detection of a sinusoidal signal in a fluctuating noise masker [2]. The crucial feature for achieving this enhancement was that the fluctuations should be coherent or correlated across different frequency bands. They called this across-frequency coherence in their demonstrations ``co-modulation.'' Therefore, the enhancement in signal detection obtained using coherent fluctuation, i.e., this phenomenon of reduced masking threshold, was called ``Co-modulation Masking Release (CMR)''. Many psychoacoustical experiments were carried out [3,4,5] and the same phenomenon was demonstrated repeatedly. The condition when CMR can occur was revealed, but a less computational model using this condition was proposed.
On the other hand, we have been tackling the problem of segregating the desired signal from noisy signal based on auditory scene analysis (ASA) [6]. We stress the need to consider not only the amplitude spectrum but also the phase spectrum when attempting to completely extract the signal from a noise-added signal which both exist in the same frequency region [7]; based on this stance, we seek to solve the problem of segregating two acoustic sources -- the basic problem of acoustic source segregation using regularities (ii) and (iv) of the following regularities: (i) common onset and offset; (ii) gradualness of change; (iii) harmonicity; and (iv) changes occurring in an acoustic event [6].
This paper proposes a computational model of CMR that consists of two models, our auditory segregation model and the power spectrum model of masking proposed by Patterson et al., and a selection process.