Conclusions

This paper proposed a method of signal extraction from noise-added signal using the amplitude envelope, the output phase and the input phase obtained from a noise-added signal passed through the wavelet filterbank. Results of simulation using this method show that a sinusoidal signal can be easily extracted if it is added to AM bandpassed noise, and that it cannot be extracted if it is added to a bandpassed noise. In applications, we feel that any signal can be segregated from a noisy signal. However, because it is impossible to segregate for a signal without using certain constraints, the proposed model extracts desired signal from a noisy signal using constraints based on auditory scene analysis.

If the parameters of wavelet filterbank are set to human auditory properties, it can be shown that the masking of a sinusoidal signal can be released as a function of the bandwidth of bandpassed noise when noise is AM bandpassed noise, and that it can not be released as a function of the bandwidth when noise is a bandpassed random noise. From these results, it can be interpreted that the proposed model is a computational model of co-modulation masking release.

In this paper two of the four heuristic regularities proposed by Bregman, (2) gradualness of change and (4) changes in an acoustic event, were considered as physical constraints in order to determine the input phase. Future work includes determining the input phases $\theta_{1k}(t)$ and $\theta_{2k}(t)$. We feel that these parameters can be determined using physical constraints related the remain regularities, (1) common onset and offset and (3) harmonicity. We also feel that the proposed model can extract the desired complex tone and speech from noisy signals using the above constraints.