Speaker - Independent V O M L Recognition : Comparison of Backpropagation and Trained Classification Trees

A series of experiments compare performance of trained classification trees t o multi-layer feedforward netRecent advances in classification with neural networks works on speaker-independent vowel recognition using inforhave been paralleled by advances in other techniques, and mation in a single spectral slice. The vowel stimuli are it is important to examine the most promising of these. In exemplars of 12 monophthongal vowels of American English this paper, we compare the performance of backpropagation taken from all phonetic contexts in spoken utterances. The (BP) networks and trained classification trees (CT) on an training set consists of 342 vowel tokens provided by 320 important real-world problem speaker-independent speakers, and the test set consists of 137 tokens provided by classification of vowel sounds in natural continuous speech. a different 100 speakers. The classification trees and neural The mathematical details of the CT technique are classifiers are trained and tested on identical data. In addipresented in a book by ~ ~ ~ i ~ ~ ~ , ~ ~ i ~ d ~ ~ , olshen, and tion, experiments are performed t o determine the most Stone [4]. Given training da ta in the form of a set of muleffective way t o present vowel information for classification. tidimensional vectors { l, the top node of the binary Classification performance is compared using (a) spectral classification tree is built by choosing a threshold for one of coefficients from the DFT; (b) Spectral coeflkients from the the variables of { x }. This threshold is chosen to maximize pitch-synchronous DFT (PS-DFT); (c) features describing the class separability of the data which passes down the the six largest peaks in the spectrum; and (d) features two descendent branches of the top node. The input space derived from Principal component analysis of the spectra, is then divided for the two descendent nodes and each of using an unsupervised neural network. these nodes continue t o split within a subset of the original The results show tha t neural nets trained with backinput space. These splits (and tree growth) continue recurpropagation produce better results than classification trees sively until all training da ta are assigned t o a leaf node in all comparable experimental conditions. Spectral with a unique and accurate cllabel. performance. Relative t o these results, using measurements the training data , a clever "pruning" criteria is then used, reducing the size of the tree and allowing for the best possiof spectral peaks produces inferior performance, while using ble performance outside the training set. The final tree, if principal component analysis t o eliminate redundancy in the spectrum slightly improves performance while greatly necessary, will fit arbitrary non-linear decision regions. reducing the size of the network. Possible reasons for the superior classification performance obtained by the backOne Of the attractive features of the "-Sive propagation are cornclassification tree technique is the ability to accurately utilpared t o other speaker-independent vowel classification stuize VeW different types of input variables. For example, if dies. the input patterns consist of both ordinal and categorical data , appropriate splits could still be made. Another possible advantage is the flexibility t o have detailed fits in part of the input space while maintaining much smoother fits in