Collection of phoneme samples using time alignment and spectral stationarity of speech signals

An automatic method for collecting a large number of phoneme samples to be used as training data for speech recognition is described. Time alignment and spectral stationarity of speech signals are used to transfer phoneme labels from a hand labeled utterance of a standard speaker to a similar utterance of another speaker for whom training data are needed. Experimental results based on speech data obtained from eight male speakers show that automatically obtained training data almost yield the same phoneme recognition accuracy as hand labeled training data.