Low-Power Convolutional Recurrent Neural Network For Monaural Speech Enhancement

Speech enhancement (SE) algorithms based on deep neural networks have achieved encouraging performance. How-ever, computationally intensive and time consuming involved in these methods. For example, the state-of-the-art deep complex convolutional recurrent network (DCCRN) requires 15G floating point operations (FLOPs) to process a one-second audio. In previous research, we proposed a convolutional recurrent network leveraging short-time discrete cosine transform (DCTCRN), which outperformed DCCRN. And the complexity of DCTCRN is only one-third that of DCCRN. However, it is still difficult to deploy DCTCRN on devices with limited hardware resources or in applications with strict latency requirements. To address these problems, we further reduce the computational complexity of DCTCRN in this work. Experimental results show that training a randomly initialized model, which contains convolutional re-current neural network and dense layers (DCTCLD) using DCT as input features, achieves competitive performance. DCTCLD's FLOPs are only 70M to handle a one-second speech, which is about 1% of DCTCRN. It is more suitable for low-power applications at the cost of certain performance degradation compared with DCTCRN. Moreover, the performance of DCTCLD is comparable to that of the DCTCRN model which costs 160M FLOPs for a one-second clip.