Development of isolated Malay words speech recognition prototype using recurrent neural network / Irdhan Sarbini

In this research, a prototype of isolated Malay word speech recognition using
recurrent neural network (RNN) is proposed. The research is working on speaker
independent, which is combination of male and female respondent. A simple three-layer
RNN which is Elman Network is employed to learn the pattern of speech features. Melfrequency
Cepstral Coefficient (MFCC) feature is selected and the features are extracted
by using Speech Filing System freeware application. Experiments are performed to
determine the optimal number of hidden neurons for the architecture of RNN. The total
recognition rate is 95 %. This research also reveals that RNN is able to give good
performance for speech recognition and for incomplete data.