On the Use of Frame-level Information Cues for Minimum Phone Error Training of Acoustic Models

This paper considers discriminative training of acoustic models for Mandarin large vocabulary continuous speech recognition. Two frame-level information cues were explored and integrated into the minimum phone error (MPE) training. First, the frame-level entropy of Gaussian posterior probabilities obtained from the word lattice of the training utterance was exploited to weight the framelevel statistics of the MPE training. The purpose of using entropy is to further emphasize or deemphasize the associated training statistics of plausibly correct and competing models for better discrimination. Second, we presented a new phone accuracy function based on the frame-level accuracy of hypothesized phone arcs instead of using the raw phone accuracy function of the MPE training. The underlying characteristics of the presented approaches were extensively investigated and their performance was verified by comparison with the original MPE training approach. Experiments conducted on the broadcast news collected in Taiwan showed that the presented approaches could achieve slight but consistent improvements over the baseline system.