Optimization of Units for Continuous-Digit Recognition Task

摘要

The choice of units, sub-word, is generally based on the size of the vocabulary and the maount of training data. In this work, we have introduced new constraints on the units: 1) they should contain sufficient statistics of the features and 2) they should contain sufficient statistics of the vocabulary. This led to minimization of two cost functions, first based o nthe confusion between the features and the units and the second based on the confusion between the units and the words. We minimized first cost function by forming broad phone classes that were less confusing among themselves than the phones. The second cost function was minimized by coding the wond-specific phone sequences. On the continuous digit recgnition task, the broad classes performed worse than the phones. The word-specific phone securences however significantly improved the performance over both the phones and the whole-word units. In this paper we ciscuss the new constraints, our specific implementation of the cost functions, and the corresponding recognition performance.