Phoneme Recognition using Competitive Neural Trees

This paper applies the Competitive Neural Tree (CNeT) method to phoneme recognition, a pattern classiication problem. CNeTs combine the advantages of Decision Trees and Competitive Neural Networks. The CNeT algorithm works by hierarchically clustering given examples while growing a tree. Diierent search methods, as well as stopping and splitting criteria are discussed. The CNeT algorithm allows to use probability estimation and interpolation for the recall. Using a phoneme recognition task as benchmark, the CNeT performs as well as a MLP, but has a much lower time complexity. Phoneme recognition is one of the earlier steps in automated speech recognition. First the speech signal is recorded and digitized. The next step is feature extraction. The raw sample is divided into time frames and for each frame a characteristic set of coeecients { a feature vector { is computed. The phoneme recognition component processes a sequence of feature vectors to derive a pho-netic labeling of the speech signal. From the phoneme sequence more abstract representations, such as sequences of syllables, words, and sentences, are derived in subsequent steps. The phoneme recognition task consists in producing from a given sequence of feature vectors a phoneme label or a probability estimation of membership in a phoneme class. In this paper we introduce a phoneme recognition component that uses Competitive Neural Trees (CNeT). We compare it to a system using Multi-Layer Perceptrons (MLP) trained with the well known Back-Propagation algorithm and with a system using Nearest Neighbor (NN) classiication. The CNeT has a structured architecture. A hierarchy of identical nodes forms an m-ary tree. Each node contains m slots s 1 ; s 2 ; : : : ; s m and a counter age that is incremented each time an example is presented to the node. Each slot s i stores a prototype p i , a counter count, class counters count j , and a pointer to a node. The prototypes p i , that have the same length as the input vectors x, are trained to become centroids of example clusters R i. The class counters count j and count are incremented each time the prototype of the slot is updated for an example of class C j. A child node may be assigned to the slot via the pointer.