A large phonemic time-delay neural network technique for all Mandarin consonants recognition

A large phonemic time-delay neural network (LPTDNN) technique has been developed to recognize all Mandarin consonants of the entire set of Chinese syllables. By modular construction, six component networks, viz. (b,d,g,ot), (l,m,n,r,ot), (j,z,zh,ot), (c,q,ch,ot), (f,s,x,h,sh,ot) and (p,t,k,ot), are formed to represent all 21 Mandarin consonants (where ot=others). The technique was thoroughly tested with 8 sets of 1157 isolated Hanyu Pinyin syllables, with 6 sets used for training and 2 sets used for testing. The overall result shows a high recognition performance with 99.2% for the inside test (i.e. testing on the datasets which were used entirely in training) and 92.7% for the outside test (i.e. testing on the remaining datasets which were never used in training).<>