A Computation-Efficient Neural Network for VAD using Multi-Channel Feature

Voice activity detection serves as an essential pre-processor in modern speech processing systems. It classifies audio segments into speech and nonspeech. Many state-of-the-art meth-ods have been proposed to increase the detection accuracy. How-ever, there are still significant limitations to retaining high per-formance while keeping low computation complexity, especially in handling unseen noises. This paper proposes a computation-efficient neural network using a multi-channel audio feature. The audio feature is contextual-aware with positional information and is represented in a three-channel way, similar to RGB pictures, which enables convolutional kernels to capture more information simultaneously. Meanwhile, we introduce channel attention inverted blocks to build a computation-efficient neural network. Our proposed method shows superior performance with extremely few floating point operations as compared with baseline methods.