Speech control feature extraction for intelligent robotics under noisy environments

Despite of significant progress on speech recognition, current techniques cannot satisfy the demands of real applications in robot controls, the main reason is that various noises in environments of robot control substantially degrade the performance of speech recognition. A feature extraction method is proposed based on sparse coding. This method makes use of the de-noising merit of sparse coding and extracts features after removing noise in Mel-frequency domain. Such a strategy integrates spare coding into speech feature extraction and can reduce the effect of noise. Experiments in speech recognition tasks show that the feature proposed possesses strong robustness against various noises and improves the performance of speech recognition in noisy environments.