Deep neural network based acoustic model using speaker-class information for short time utterance

In speech recognition, it is preferable not to hypothesize the details, e.g., specific age and gender, of a target user. However, speaker independence is one of the things that degrades ASR performance. In this work, we propose a speaker adaptation method to recognize a short time utterance. There have been several studies on speaker-independent DNN-HMM in which i-vector is computed, and the additional information is combined with acoustic features. However, it is difficult to calculate i-vector accurately or apply speaker adaptation (e.g. fMLLR) when the utterance time is short (0.5sec~). In our approach, we calculate the similarity score between the speaker class and the target utterance and utilize speaker class information configured in advance. As a precondition, we restrict the available time period to the first 50 frames per utterance for the recognition of short utterances. In experimental tests, we obtained a 4.0% relative WER gain compared to conventional DNN-HMM.