Inter-Phone and Inter-Word Distances for Confusability Prediction in Speech Recognition

In this work we investigate new inter-phone and inter-word distances and we apply them to predict if two words of the lexicon of an Automatic Speech Recognition (ASR) system are likely to be confused. The inter-word distance is calculated from an alignment between the phonetic transcriptions of the words by adding the distances between the aligned phones. We bring a new solution in which the inter-phone distance used for computing the inter-word distance is not the same used to compute the phonetic alignment. The first one is calculated between the acoustic models of the phones with a new formula that we propose. The second one is based on phonetic knowledge. We also use two different kinds of alignments: either with or without insertions and deletions. In order to evaluate the performances, we introduce a classical false acceptance/false rejection framework and the prediction Equal Error Rate (EER) was measured to be less than 2%.