The Research And Application Of Voice Wake-up With Deep Learning

Voice wake-up is a special speech recognition technology,which is applied to various intelligent devices with the advancing of mobile internet and artificial intelligence.Voice wake-up plays a key role in opening the mobile assistants,vehicle environments or smart home environments.Voice wake-up technology has been developed forward,but it still faces some problems in practical application scenarios where the recognition accuracy is poor in noisy and far-field environments.In some platforms with low computing performance,it also faces the problem of relatively high computational complexity and resource consumption.This paper focuses on the above issues,optimizes the acoustic model,applies a decoding algorithm with relatively low computational complexity,and is dedicated to improve the performance of the voice wake-up system.The developed system is also applied in actual projects.The main works of this paper includes:1.In order to improve the accuracy of the wake-up system in the noise and far-field environments,the speech crpus is enlarged with noise and simulated far-field data.A multi-structured and streamlined voice wake-up acoustic model is trained,and the Viterbi algorithm is used for path search.Finally,this paper realize a HMM/Filler based voice wake-up system2.Based on the confidence of decoding calculation,this paper also realizes the voice wake-up system of dedicated wakeup words and customizable wake-up words.After experimental demonstration,the confidence-based voice wake-up system can achieve better recognition performance compared with the HMM/Filler based speech wake-up system.3.Engineering application of the voice wake-up system.The algorithm of the voice wake-up system of the voice is transplanted to the platform of Android,and the corresponding SDK is designed to verify the feasibility of the voice wake-up system on the mobile terminal.