Connectionist acoustic word models

Abstract

Other researchers have claimed significant improvements to their recognizers by using word models based on data-driven subphonetic units rather than traditional subword models. A possible advantage of this approach is that subphonetic models can be derived automatically from the data, so that the recognizer is trained to discriminate between acoustic categories. The authors describe some of the problems with the units that are derived from acoustic-phonetic considerations (when used for a hidden-Markov-model-based recognizer), and propose a novel technique for constructing acoustic word models using a multilayer perceptron (MLP). The authors are designing a subphonetic unit called the UNnone which is similar to fenones. A vector quantizer is used to partition the acoustic space into a set of clusters. Once the vector quantizer has been designed, the training vectors are compared to the reference vectors using a Euclidean distance measure. The label corresponding to the closest reference vector is assigned to the input vector. These labels are used as targets for training the MLP.<>