Digital Signal Processing Reference
In-Depth Information
Fig. 5.8 shows phoneme classification as a function of the average length of
the phoneme. It can be observed that for
MLP
1
0
, the phoneme classification
is quite independent of the length of a phoneme. On the other hand, for the
remaining MLPs, it can be seen that there is a high dependency with the
length. For long phonemes, the common label equals the current label and
the classification performance increases.
Hence, it can be concluded that a relative good performance of a particular
MLP is achieved as long as the common and current label match. In addition,
it is highly probable that a MLP misrecognizes a phoneme to the most similar
phoneme, i.e. that one which shares a high number of articulatory attributes.
Therefore, in the example presented in Fig. 5.6, if the
nasality
phone [ ae]
is presented, it is more probable that the MLP confuses it with e.g., /eh/
rather than /n/, estimating thus lower probability for /n/. This fact is based
on the discriminative criterion that the MLP learns based on detecting the
most similar articulatory attributes.
a
b
c
Phoneme
Labels
Cepstral
Featur es
MLP
1
−
2
MLP
1
−
1
Featur e
Level
M
=5
MLP
1
0
MLP
1
1
MLP
1
2
Common
Label
bb
bb
b
Intermediate
Posteriors
Posterior
Level
MLP
2
Final
Posteriors
Fig. 5.9.
Modifying training process for stressing intra-phonetic information at the
output of the first hierarchical level. Each MLP is trained based on the
common
label.
Nevertheless, as it was mentioned before, the
nasality
attribute of the phone
[ae] preceding the phoneme /n/ may be useful for a better estimation of /n/
in the hierarchical framework. This attribute is by now blurred, due the pri-
ority that the MLP gives to recognize the current label (/ae/). Therefore, in
order that the MLP concentrates on the nasality attribute, a modification from
the current to the common label is implemented in the training process as it is
Search WWH ::
Custom Search