The Research On Voice Wakeup Technology Based On Transfer Learning

Voice Wakeup is mainly used on the devices such as smart speakers,and it's a new entrance for human-computer interaction,which makes it become an important research direction of intelligent speech.Due to the existence of noise and far recognition in our life,a lot of challenges to Voice Wakeup need to be solved.In addition,the performance requirements of hardware devices for memory size,computing resources and power,also require improvement of Voice Wakeup.In order to solve these problems,we study the Voice Wakeup system based on deep supervector and transfer learning,aiming to improve the performance of it.This work includes the following parts:1.The effect of Voice Wakeup was improved by optimizing the parameters of the DNN-HMM model,such as the number of input frames and model dimensions.The noise and far-field data were added to the clean data,in order to promote the system performance from the source of the data.Finally,the sparsity was used to improve training speed.2.For the registration and identification of Voice Wakeup,a new registration and testing method based on Deep Supervector was proposed under the improved DTW method.Besides,the Cosine distance was used to measure the matching degree of the template.3.The Voice Wakeup system based on transfer learning was proposed.The knowledge of the teacher model was used to guide the training of the student model to improve the recognition effect of Voice Wakeup,which is mainly realized by the soft tag.4.The Voice Wakeup system was implemented in the ARM platform.The voice activity detection(VAD)was used,and two functions of awakening and command word recognition realized.