Speaking rate dependent multiple acoustic models using continuous frame rate normalization

This paper proposes a method using speaking rate dependent multiple acoustic models for speech recognition. In this method, multiple acoustic models with various speaking rates are generated. Among them, the optimal acoustic model relevant to the speaking rate of test data is selected and used in recognition. To simulate the various speaking rates for the multiple acoustic models, we use the variable frame shift size considering the speaking rate of each utterance instead of applying a flat frame shift size to all training utterances. The continuous frame rate normalization (CFRN) is applied to each of training utterances to control the frame shift size. Experimental results show that the proposed method outperforms both the baseline and the conventional CFRN on test utterances.