Digital Signal Processing Reference
In-Depth Information
variance of the signal to noise ratio is used as the address index to a specific
MCMAC memory location and the third dimension of the MCMAC memory
is the frequency.
The ASE algorithm as configured in Figure 8-5 operates in the frequency-
domain since we need to compute the spectral value used in SNR
computation (8-9a, 9b). Using the learning rule as expressed in (3-6), we
compute the profile of the Weiner filter weights and store them in the
MCMAC memory. These new or updated values from the MCMAC memory
are read and used as the Weiner filter weights in estimating the enhanced
speech without any musical noise artefacts. Subsequently, we use the
enhanced speech to calculate the SNR values needed in the ASE module and
Figure 8-5.
Block Diagram of the Proposed ASE-MCMAC Using a Microphone Pair.
the SNR variance module. Therefore, this closed-loop nature of the ASE-
MCMAC allows the system to employ an unsupervised learning.
In the recall mode, however, the information needed for the computation
of address indices are all in the time-domain, except for the processing of the
noisy signal by the Wiener filter. We could have carried out this step in the
frequency-domain as well, but that would have required a number of FFT
