Speaker Recognition Based on Lightweight Neural Network for Smart Home Solutions

With the technological advancement of smart home devices, the lifestyles of people have been gradually changed. Meanwhile, speaker recognition is available in almost all smart home devices. Currently, the mainstream speaker recognition service is provided by a very deep neural network which trained on the cloud server. However, these deep neural networks are not suitable for deployment and operation on smart home devices. Aiming at this problem, in this paper, we propose a packet bottleneck method to improve SqueezeNet which has been widely used in the speaker recognition task. In the meantime, a lightweight structure named TrimNet has been designed. Besides, a model updating strategy based on transfer learning has been adopted to avoid model deteriorates due to the cold speech. The experimental results demonstrate that the proposed lightweight structure TrimNet is superior to SqueezeNet in classification accuracy, structural parameter quantity, and calculation amount. Moreover, the model updating method can increase the recognition rate of cold speech without damaging the recognition rate of other speakers.