Interspeech'2005 - Eurospeech
This paper wishes to contribute to the solution of the problem occurring when an automatic speech recognition system does not recognize an input utterance correctly. The solution is usually based on a utilization of a confidence measure (CM) which is assigned to each recognized word and which informs a user or a higher level module on the belief that the recognized word has been really said. The task becomes more difficult if the vocabulary contains acoustically similar words which differ for example only in one phoneme. To cope with this problem, we introduced a new confidence measure based on our previous experiments. The basic elementary unit for which the presented CM is investigated is a phone.
The first part of the article shortly describes the used speech recognition system and the previously used confidence measures. The main part of this article deals with description of creation of the new CM based on utilization of artificially created training data and also with description of the used classification features and the classifiers based on this CM.
A rejection technique based on the described CM was evaluated on the Czech yellow-pages database. Experimental results show that the proposed rejection technique achieves approximately 5% equal error rate (ERR) for phone rejection and about 6-16% EER for word rejection.
Bibliographic reference. Bartos, Tomás / Müller, Ludek (2005): "Detection of recognition errors based on classifiers trained on artificially created data", In INTERSPEECH-2005, 3361-3364.