Abstract

this paper, we investigate the use of confidence measures for the evaluation of pronunciation models. The confidence measures and pronunciation models are obtained from the ABBOT hybrid Hidden Markov Model/Artificial Neural Network (HMM/ANN) Large Vocabulary Continuous Speech Recognition (LVCSR) system [4] and the experiments were carried out using the North American Business News (NAB) and ARPA Hub4 Broadcast News (BN) corpora. 2. CONFIDENCE MEASURES AND PRONUNCIATION MODELS A confidence measure may be defined as a function which quantifies how well a model matches some acoustic data. More specifically, an acoustic confidence measure is one which is derived exclusively from an acoustic model. A pronunciation model for a word specifies how it is believed to be articulated in terms of a sequence of subword acoustic classes. The goal when evaluating a pronunciation model is for some word is to determine how well the model matches acoustic realisations of that word. Therefore, an acoustic confidence measure is naturally suited to the task. A common approach to evaluating pronunciation models, however, is to align the subword class sequence output by the recogniser, using full word level decoding constraints, against an alternative subword sequence obtained without any pronunciation model constraints. In this case, a poor pronunciation model is signalled by a portion of the alignment where the class labels do not match. This approach is undesirable for two reasons. Firstly, the alignment only signals pronunciation variants and does not give a direct measure of model match and secondly, obtaining an accurate alternative decoding sequence is difficult. One method for obtaining such a decoding sequence is to transcribe the acoustic data with subword class labels by hand, e.g. ...

Citations