Different confidence measures for word verification in speech recognition

Abstract

Recent research in Automatic Speech Recognition (ASR) technologies has shown the key-word spotting (KWS) systems as one of the most interesting options for accessing information using speech. KWS systems can accept spontaneous speech, which allows potential users to ask for information without learning complex protocols for the human–machine communication. One of the most relevant aspects in KWS systems is the verification of key-word candidates. Utterances detected as key-words could be either `false alarms' (non-key-words or incorrectly recognized key-words) or `correct key-words'. The use of confidence measurements allows (by additional processing of the spoken sentence) the verification of the candidates and the decision as to whether each utterance must be accepted as a correctly recognized key-word or rejected as a false alarm. In this work we propose a novel method for verification in those KWS systems based on phone models. Under our new approach, a phonematic speech recognizer decodes the spoken sentence in parallel with the KWS recognizer. The first one produces a phone string as output while the second one generates a key-word/filler-model string. By aligning both strings, a set of characteristics is extracted which are used to verify the putatives key-word. For that we have built two classifiers; in the first one the euclidean metric is modified and adapted in a local and iterative way in order to give greater importance to the most discriminate directions between the classes. The second is a vector quantizer which was trained using adaptative technique learning. We have applied the proposed method to several KWS tasks. Experimental results presented in this paper show that the proposed verification method improves the performance of the KWS systems by reducing the false alarm rate without a significant increase in the rejection of correctly detected key-words.