Digital Signal Processing Reference
In-Depth Information
the lower frequency channel because the spectra of stationary car noise is
concentrated in the lower frequency region. Thus,
R
is a vector with 84
elements. As shown in Figure 19-1, t h e microphone is the one nearest
to the driver. Finally, the 84 elements are normalized so that their mean
and variance across elements are 0 and 1.0, respectively. Prototypes of
noise clusters are obtained by applying the k-means algorithm to the
feature vectors extracted from the training set of noise signals.
An example of the clustering results are illustrated in Table 19-3, where
we how many samples of each driving condition each noise class contains
when four clusters of noise are learned. As seen from the table, clus-
ters are naturally formed for 'normal', 'music playing', 'fan' and 'open
window' situations, regardless of the driving speeds. From the results,
it is expected that the relative power of the sound signals at different
microphone positions can be a good cue for controlling MRLS weights.
4. IN-CAR SPEECH CORPUS FOR
DISTRIBUTED MICROPHONE
The distributed microphone speech corpus is a part of the CIAIR (Cen-
ter for Integrated Acoustic Information Research) in-car speech database
