Word Confusability as an Evaluation Criterion of the Hardness of a Speech Recognition Task

Evaluation of hardness of a given automatic speech recognition task - once a specific language has been chosen - is usually accomplished in terms of such factors as the vocabulary size, the complexity of the grammar, isolated vs. continuous speech, and the acoustic conditions. In practice, little attention has been paid so far to the fact that, once a vocabulary has been fixed, along with a given grammar and setup of environmental conditions, the difficulty of the recognition task may vry according to the amount of 'confusable' words in the vocabylary, as well as to the specific 'degree of confusability' among them. This report informally introduces the concept of confusability, briefly surveying technical literature where the problem is faced. It then introduces two formal definitions of the problem, the former having a statistical root in Bayes' decision theory, and the latter relying on a geometric approach, involving the concept of 'distance' between pairs of words. Practical hints on possible directions to follow in order to apply the definitions are finally outlined