A comparative study of OCON and MLP architectures for phoneme recognition

In this paper a comparative study between One-Class-OneNetwork (OCON) and Multi-Layered Perceptron (MLP) neural networks for vowel phoneme recognition is presented. The OCON architecture, first proposed by I.C.Jou et al 1991, is similar in design to a conventional feed-forward MLP, only each class had its own dedicated sub-network containing a single output node. Conventional MLPs usually consist of fullyconnected nodes which not only result in a large number of weighted connections but also create the problem of cross-class interference. Using vowel phoneme data from the DARPA TIMIT corpus of read speech, MLP and OCON architectures were trained and the relative effects of recognition and convergence rates during both intra and inter-class adaptation tested. The OCON showed an increase in the convergence rate of 273% and an improvement of adapted recognition rates against the MLP of over 12%. However, due to the isolated nature of each OCON class, it was unable to utilise inter-class information. This resulted in a recognition rate reduction of over 6% for unadapted phonemes during adaptation of remaining vowels, compared with the MLP results.