One-Sample Speech Recognition of Mandarin Monosyllables using Unsupervised Learning

In the speech recognition, a mandarin syllable wave is compressed into a matrix of linear predict coding cepstra (LPCC), i.e., a matrix of LPCC represents a mandarin syllable. We use the Bayes decision rule on the matrix to identify a mandarin syllable. Suppose that there are K different mandarin syllables, i.e., K classes. In the pattern classification problem, it is known that the Bayes decision rule, which separates K classes, gives a minimum probability of misclassification. In this study, a set of unknown syllables is used to learn all unknown parameters (means and variances) for each class. At the same time, in each class, we need one known sample (syllable) to identify its own means and variances among K classes. Finally, the Bayes decision rule classifies the set of unknown syllables and input unknown syllables. It is an one-sample speech recognition. This classifier can adapt itself to a better decision rule by making use of new unknown input syllables while the recognition system is put in use. In the speech experiment using unsupervised learning to find the unknown parameters, the digit recognition rate is improved by 22%.