Active learning: theory and applications to automatic speech recognition

摘要: We are interested in the problem of adaptive learning context automatic speech recognition (ASR). In this paper, we propose an active algorithm for ASR. Automatic systems trained using human supervision to provide transcriptions utterances. The goal Active Learning is minimize training acoustic and language models maximize performance given transcribed untranscribed data. aims at reducing number examples be labeled by automatically processing unlabeled examples, then selecting most informative ones with respect a cost function label. paper describe how estimate confidence score each utterance through on-line lattice output recognizer. scores filtered informativeness optimal subset samples selected. has been applied both batch scheme have experimented different selective sampling algorithms. Our experiments show that amount data needed word accuracy can reduced more than 60% random sampling.