Application And Implementation Of Voice Wake-up Technology In Voice Assistant System

With the rapid discovery of Internet technology and communication technology,people have been brought into the era of global informatization,and computers and mobile phones have become important tools for people to work,communicate and entertain.In the context of the era of artificial intelligence,AI voice has also been greatly developed,and voice assistants have become an indispensable software for intelligent terminal devices,which has put people on the intelligence road of the hands "free".Voice recognition is the key technology for humans to interact with computers,and voice wake-up is the entrance to voice recognition and the primary condition for humans to control smart devices from a distance.The research work on voice wake-up has only begun in recent years.Compared with the research work on speech recognition,it is much less mature in technology.This paper studies the application and implementation of the voice wake-up technology on the voice assistant system,and implements the application of the wake-up word "Little T Little T" in the voice assistant.The main work is as follows:1.The core algorithm of voice wake-up technology is studied.The voice wake-up task is a small resource-level keyword retrieval task.The current commonly used system structure is the Keyword / Filler system and end-to-end system based on the hidden Markov model.Discuss the end-to-end system,that is,the input is the voice characteristics,and the output is the keyword results.In this paper,an end-to-end keyword retrieval system based on the attention mechanism is adopted.Experiments are conducted on the encoder structure,attention mechanism,and the addition of convolutional layers.The best speech wake-up neural network model is selected through comprehensive comparison.2.Noise and reverberation are factors that affect the wake-up of speech.For this,this paper studies the neural network-based speech enhancement,and conducts an experimental study with a generative confrontation network model,comparing three training methods: single denoising,single dereverberation,As well as the joint training method of denoising and dereverberation,experiments have shown that the combined training method of dereverberation and denoising is better than the other two methods in achieving better robustness of speech wake-up.3.Design and implement a voice assistant system through functional requirements analysis,data analysis,and process modeling of the voice assistant system.The system mainly includes a voice wake-up module and an intelligent dialogue module.The intelligent dialogue module includes three sub-modules: speech recognition module(ASR),natural language processing module(NLP),and speech synthesis module(TTS).The composition of each module is analyzed in this paper.And data flow,and designed the data structure,completed the construction of the voice assistant system.This paper introduces the application architecture of voice wake-up technology on the voice assistant system.It has been proved by experiments that the application of voice wake-up technology is realized.Through the verification of the application of voice wake-up technology in the voice assistant system,multi-scenario tests have been conducted on smart terminal devices(mobile phone and TV)to basically complete the task of voice wake-up.Continuous stress testing on smart TV terminals can generate 1000 data in about 1 hour,and eventually reached a wake-up rate of 95.58%,with false wake-up of approximately 0.6 times / 24 hours.