Comparison of classic and hybrid HMM approaches to speech recognition over telephone lines

The subject of the present dissertation is the automatic speaker-inde¬ pendent recognition of isolated German digits spoken (by Swiss people) over the public switched telephone network. The approaches considered for this task are all based on hidden Markov models (HMMs). In ad¬ dition to the classic HMM approaches, several connectionist ideas are investigated in order to improve the discrimination capability of the classic HMM systems. In the first part, the theoretical foundations of HMMs in general and discrete density HMMs (DDHMMs) in particular are summarized. After that, several connectionist ideas, also termed hybrid HMM sys¬ tems in the sequel, are theoretically discussed. In particular, a socalled connectionist-SCHMM approach based on classic semi-continuous HMMs (SCHMMs), which has been proposed by the author, is intro¬ duced. The second part goes into the details of the design of the experimen¬ tal system RECO, which is based on DDHMMs. It covers the speech data collection for the training and test data as well as the different ex¬ periments that have led to the high performance of the recognizer. With 98.6 % word recognition rate, the resultant DDHMM recognizer yields the same performance as the COST 232 reference recognizer based on continuous density HMMs on the same speaker-independent test set. Different connectionist ideas of the first part are compared with each other and with classic HMM approaches in the last part of the thesis. The comparison shows, on the one hand, that the connectionistSCHMM system proposed exhibits the best performance of all hybrid