"\"Peak\" Quantization: A Training Method Suitable for Terminal Equipment to Deploy Keyword Spotting Network".

In order to reduce the storage space occupied by the well-trained Keyword Spotting (KWS) network model during deployment on terminal devices while maintaining the recognition accuracy of the model as much as possible, a new method of 'peak' quantization is proposed. By limiting the maximum value of the network model weights to change the quantization process, it has achieved good results in reducing the loss of recognition accuracy. Compared with other Network quantization, such as traditional quantization and Alternating Direction Method of Multipliers (ADMM) quantization, it shows the advantages and characteristics of 'peak' quantization. Through related experiments with the Google voice dataset and the MNIST dataset, its universality is demonstrated.