Training data selection for improving discriminative training of acoustic models

摘要

This paper considers training data selection for discriminative training of acoustic models for large vocabulary continuous speech recognition (LVCSR). Three novel data selection approaches are proposed. First, the average phone accuracy over all hypothesized word sequences in the word lattice of a training utterance is utilized for utterance-level data selection. Second, phone-level data selection based on the difference between the expected accuracy of a phone arc and the average phone accuracy of the word lattice is investigated. Finally, frame-level data selection based on the normalized frame-level entropy of Gaussian posterior probabilities obtained from the word lattice is explored. The underlying characteristics of the presented approaches are extensively investigated and their performance is verified by comparison with standard discriminative training approaches. Experiments conducted on a broadcast news speech transcription task show that with the aid of phone- and frame-level data selection we can reduce more than half of the turnaround time for acoustic model training and simultaneously obtain a comparably good set of discriminative acoustic models.