Learning speed improvement using multi-GPUs on DNN-based acoustic model training in Korean intelligent personal assistant

Abstract

This paper proposes a learning speed improvement using multi-GPUs on DNN-based acoustic model training in Korean intelligent personal assistant (IPA). DNN learning involves iterative, stochastic parameter updates. These updates depend on the previous updates. The proposed method provides a distributed computing for DNN learning. DNN-based acoustic models are trained by using 320 h length Korean speech corpus. It was shown that the learning speed becomes five times faster on this implementation while maintaining speech recognition rate.