Abstract
This paper describes experiments on recognition of vowels in 212 Japanese words. Recognition is based on frequency distribution of spectral local peaks for every frame. The spectrum obtained from a band pass filter bank sampled every 10 ms is represented by a binary vector, each element denoting the presence or absence of a local peak at a particular channel. The conditional probability distribution of local peaks and the joint probability between one local peak and another are obtained by a statistical analysis. Recognition procedures are derived from the functional form of these underlying probability distributions. The power information in certain channels is also used as a supplementary one for improving the recognition accuracies. The recognition rate is 90.8% for designed samples, and 85.3% for test samples uttered by5males. The effects in transfer systems with various frequency characteristics on vowel recognition are investigated. The experimental results indicate that the effects are significant, and the proposed method using local peaks is robust for the different frequency characteristics studied.
