Automatic Speech Recognition in the presence ofadditive background noise is a challenging task. The'missing data' approach to this problem relies onidentifying spectral-temporal regions which aredominated by the speech source. The remaining regionsare considered to be 'missing' and generally dealt witheither by being ignored or imputed using HiddenMarkov Models. In contrast to missing data methodsbased on HMMs, connectionist approaches open up thepossibility of making use of long-term time constraintsand making the problems of classification withincomplete data and imputing missing values interact.This paper addresses the problem of combining robustASR with missing data and pattern completion in asingle Recurrent Neural Network. We report isolateddigit recognition results on a realistic missing data case,in which the time-frequency regions which are missingare determined by local Signal-to-Noise Ratio estimates
展开▼
