Multi-Mode LSTM Network for Energy-Efficient Speech Recognition

Abstract

We newly introduce a novel processing scenario of long short-term memory (LSTM) network for the energy-efficient speech recognition. Compared to the conventional single-mode processing based on the fixed computing scheme, the proposed LSTM processing contains multiple operating cells providing attractive tradeoff between the recognition accuracy and the energy consumption. For the case study, the state-of-the-art LSTM network is modified to have two types of processing cells, strong and weak cells, which are dedicated to the accuracy-aware and energy-aware LSTM sequences, respectively. By allocating as many weak cells with low energy as possible, experimental results show that the proposed work saves the energy consumption for speech recognition by 75% compared to the original network.