Experimental Results
We present the rcsults of speech recognition experiments by using the HMM and the fuzzy HMM，and the rcsults of speakcr recognition experiments by using the GMM and the fuzzy，GMM.Speech data corpora used for these experiments are the TI46（Texas Instruments）and the ANDOSL（Austraiian National Database of Spoken Language）.For the TI46，there are 16 speakers，8 femaje and 8 male and the vocabulary contains 46 words which are ten digits from 0 to 9，26 letters from a to z，ten commands：“enter，erase，go，help，no，rubout，repeat，stop，start，yes”.Each speaker repeated the words 10 times in a single training session，and then again twice in each of 8 later testing sessions.For the ANDOSL，we use a subset of 108 native speakers，54 female and 54 male.Each speaker contributed in a single session，200 phonetically rich sentences，where 20 sentences are used for training，and 180 sentences are used for testing.Further details can be found from［16］to［19］.The similarity normajisation method used for speaker verificaLion is based on a posteriori probability proposed by Matsui and Furui［241.In Table 1，m=1，m=1.2.and m=1.2.δ=1.5 stand for the HMM.the FHMM，and the NC-based FHMM，respectively.The E set includes 9 letters：｛b，c，d，e，g，p，t，v，z｝.Table 2 presents results of the HMM and the FHMM.The NC-based FHMM has the same results with the FHMM in chis experiment.
Table 1 Isolated word recognition results for the E set，using 6-state left-to-right discrete HMMs，in speakerdependent mocfe（TI46）
Refer to Table 1（continue）
Table 2 Isolated word recognition results for the 10-digitset，using 6-state left-to-right discrete HMMs，in speakerdependent mode（TI46）
Table 3 presents speaker identification，where m=1，m=1.05，and m=1.05，δ=35 stand for the GMM，the FGMM，and the NC-based FGMM，respectively.Speaker verification is presented in Table 4.Speaker identification performed on 108 speakers using the ANDOSL database is in Table 5.
Table 3 Speaker identification results for 16 speakers using 10-comrnand vocabulary，in text-independent mode（TI46）
Table 4 Speaker verification results for 16 speakers，using 10-command vocabulary，in text-independent mode（TI46）
Table 5 Speaker identification results for 108 speakers，in text-independent mode（ANDOSL）
