Modeling inter-cluster and intra-cluster discrimination among triphones

Discriminative training is a major contribution to the success of automatic speech recognition (ASR) in the last decade. However, since most ASR systems employ state tying which ties `similar´ states in a cluster, discriminative training may only improve inter-cluster discrimination, but states belonging to the same cluster obviously cannot be distinguished. Recently, the concept of distinct acoustic modeling was investigated by a new acoustic modeling method called eigentriphone modeling. In the new method, states are grouped, but not tied, into separate clusters, and the difference vectors between mean vectors of the member states and their cluster center vector are modeled by a basis approach using a set of eigenvectors which are also called eigentriphones. This paper investigates whether the intercluster discrimination achieved by discriminative training and intra-cluster discrimination obtained by eigentriphone modeling are additive. In a simple procedure that is applied to each state cluster, the discriminatively trained cluster center vector is integrated with the difference vectors trained by eigentriphone modeling to construct the final mean vectors of the distinct states in the cluster. Experimental evaluation on the WSJ0 5K task shows that the two techniques are indeed additive.