Digital Signal Processing Reference
In-Depth Information
It is not difficult to see that the channel facing the speaker is
expected to contain mainly the speech from a speaker, whereas the opposite
channel will be primarily the reference signal (noise) and the
secondary and ternary echoes reflected from the windshield and the back of
the car, respectively. Understandably, there will be some portion of the
reference signal in the front channel as well due to reflection from the
chamber walls, the primary acoustic echo and the speech in the other one.
It was reported in a number of works in the automotive science that the
corruption of speech in a vehicular environment is not purely additive in
nature. In most cases, the relationship between the original speech and the
noise involves a convolution process instead of a simple addition, which has
been the norm in the Shannon-based information processing and
communication systems community. In other words, the noisy speech can be
better expressed by:
Here
s
(
t
) represents the speech input whereas,
d(t)
is the overall
degradation, which may include the impulse response of the vehicular
chamber. Since the model is not additive, Fourier analysis and the
subsequent filtering cannot be applied directly. This, in turn eliminates the
usage of the ubiquitous LMS-based ASE algorithm as it has been the case in
the majority of earlier speech enhancement techniques including our earlier
studies [4,6].
To overcome this, we have opted to resort to a high-order CMAC with a
non-linear basis function. This has allowed us to tackle both the ambient
convolutive noise term associated with the chamber and the traditional
additive noise term a-la-communication systems.
Experimental results reported later in this chapter demonstrate not only
the effectiveness of the proposed MCMAC system when coupled with an
ASE in the enhancement of the convolutive nature of noise but also the
robustness of the technique promises as a viable candidate for deployment in
the next generation vehicular communication systems.
The Signal plus Noise to Noise Ratio (SNNR) has been used as the
quantitative measure of performance in this work and it can be defined as the
sum of the
a priori
SNR and the
a posteriori
SNR values:
