BANDWIDTH EXTENSION OF TELEPHONY SPEECH - Adaptive Signal Processing: Next Generation Solutions - page 375

Figure 7.14 Typical progression of the distance between actual output and desired
output of the artificial neural network during the training phase.
When it comes to the operation mode of the artificial neural network, after the train-
ing phase has been completed, there are two major characteristics of such a network
that have to be mentioned. The first one is the low computational complexity
needed by such an artificial neural network if not too many layers or nodes are used
respectively. Such networks are able to learn complex tasks using comparatively
few layers and neurons. This is an advantage over codebook approaches as described
in the next section since the computational effort used to evaluate a distance measure
for each codebook entry is omitted using artificial neural networks. On the other hand
artificial neural networks do not offer possibilities to interfere manually the way code-
book approaches do. For example by observing the distance between the narrowband
spectral envelope of the input signal and the narrowband codebook entry producing
minimum distance one is able to predict if the bandwidth extension runs out of the
rudder completely and therefore switch off such a system.
Codebook Approaches A third approach for estimating the vocal tract trans-
fer function is based on the application of codebooks [4]. A codebook contains a repre-
sentative set of either only broadband or both bandlimited and broadband vocal tract
transfer functions. Typical codebook sizes range from N cb ¼ 32 up to N cb ¼ 1024.
The spectral envelope of the current frame is computed, for example, in terms of
N pre,nb ¼ 10 predictor coefficients, and compared to all entries of the codebook. In
case of codebook pairs the narrowband entry that is closest according to a distance
measure to the current envelope is determined and its broadband counterpart is
selected as the estimated broadband spectral envelope. If only one codebook is utilized
the search is performed directly on the broadband entries. In this case the distance
measure should weight the nonexcited frequencies much smaller than the excited