Mircea Giurgiu * Results on Automatic Speech Recognition in Romanian
Fig. 11. - Evolution
of error during the training for the four MLP: net40sw.nnw and
net50sw.nnw, respectively nvq120t.nnw and nvq140t.nnw.
Fig. 12. - Romanian spoken digit recognition results with MLP fed with the LPC spectral representation of words: net40sw.nnw, respectively net50sw.nnw.
These results (see Fig. 11 and 12) show that the recognition rate depends on the number of neurons in the hidden layer, that is, the more neurons are, the better recognition rate is. A very interesting idea has arisen during the experiments: VQLPC could be a good solution for substituting the large dimensionality of input pattern with a smaller one obtained by quantization of LPC vectors. The number of neurons from the input layer in the VQLPC approach could be 12 times less than in the LPC one. The result is that less computational time and less memory are needed at the approximately the same recognition rate.
3.3.2. MLP and Constrained Clustering Segmentation applied to speech recognition
Since MLP deals with the problem of variable word length, we have been proposed the representation of speech as a sequence of acoustically stationary frames {x1,x2,...,xT}, and the idea behind the Constrained Clustering Segmentation (CCS) technique is to locate the optimum start and stop frame numbers {(i1,j1), (i2, j2),..., (im, jm), i1=1, ik=jk-1+1, jm=T} (for m segments). The variation of the feature vectors xl within such semistationary segments is assumed to be small, so each segment (ik, jk) may be represented by a centroid ck. The centroid is defined as the parameter vector which minimizes the distortion within a sequence according to a distortion measure. The number m of segments in which an utterance is divided depends on the segmentation distortion in the CCS algorithm, on the threshold of average frame distortion, etc. [5].
Fig. 13. - The segmentation error for word
/opt=eight/, respectively word /zece=ten/ in CCS
algorithm versus the number of segments.
184
