Research On VoLTE Speech Quality Evaluation Algorithms Based On Machine Learning

With the large-scale commercialization of LTE system and the continuous development of 5G technologies,the coverage of VoLTE(Voice over LTE)is expanding with the accelerating rate of expansion,and it is possible that VoLTE will completely replace narrowband telephone networks in the future and become the mainstream communication service solution.Therefore,how to evaluate the user satisfaction level of VoLTE services flexibly and accurately has become an important issue for operators and researchers.This paper studies the quality evaluation problem of VoLTE speech.The main work and innovations are as follows:1.For the research of VoLTE speech evaluation algorithms,a VoLTE speech database based on 4G transmission network has been established,containing more than 20.000 impaired samples,with a total duration of more than 60 hours.In order to simulate various scenarios that may appear in users' actual calls,different impaired sample files are set for voice transmission,in addition to the most common communication scenario with good voice quality.The types of impairment include voice interruption,single-pass,and the change of sound decibel,etc.After the impaired transmission file is obtained,it is compared with the unimpaired sample file.The impaired transmission file is marked with the POLQA score using offline platform.2.A no-reference assessment for VoLTE speech quality algorithm is proposed,named as NAVSQ.The NAVSQ algorithm only needs the impaired transmission signal to evaluate the speech quality.44 kinds of speech feature parameters from the impaired transmission signal are extracted to quantize the VoLTE speech.The GBDT(Gradient Boosting Decision Tree)algorithm is used to construct the mapping model.Compared with the existing P.563 algorithm,the correlation between the NAVSQ's predicted results and the POLQA scores can reach 0.95.RMSE is 0.17 and MAE is 10.7%.With excellent accuracy and flexibility,the NAVSQ algorithm shows better performance than P.563.3.A Full-reference Assessment of VoLTE Speech Quality algorithm is proposed,named as FAVSQ.The main difference between this algorithm and the NAVSQ algorithm is that a variety of speech difference features are extracted from the time domain and the frequency domain,as a contradistinctive information of the reference signal and the impaired signal,and the mapping model is constructed using the GBDT algorithm.Compared with the existing full reference standard PESQ algorithm,the correlation between FAVSQ predicted results and POLQA scores are up to 0.98.RMSE is 0.57,and MAE is 9.1%,which achieves more accurate VoLTE speech evaluation.