摘要
针对语音业务质差发现难和定位难的问题,本文提出一种基于深度学习和软硬采关联的语音质量检测与优化技术,通过扩展定义新型RTP特征实现对单通、吞字和断续等异常事件的检测。引人深度学习技术构建AI MOS评估模型,得到RTP特征向量到MOS值的最优解。对标注样本DBSCAN聚类后,应用GBDT生成特征组合,通过Xgboost、LR和神经网络算法生成不同的预测值。最后基于Stacking融合,在避免过拟合的前提下得到精准MOS检测结果。在此基础上,关联软采UEMR数据,在业界首次实现对语音质差问题的无线根因定位。实际应用后,MOS检测准确率和质差点召回率均达到业界领先水平,并可输出质差优化方案,优化后小区上行质差时长占比明显下降,提升效果显著。
Voice service is one of the core services of telecommunications network,and voice quality is the key factor affecting user perception.Aiming at the problems of poor quality and difficult location of traditional voice services,this paper proposes a voice quality detection and optimization technology based on deep learning and software hardware data correlation.New RTP feature is extended to detect abnormal events such as one-way,swallow word,intermittent,etc.Deep learning technology is introduced to build an AI MOS evaluation model,and the optimal solution from RTP feature vector to MOS score is obtained.On this basis,it is the first time in the industry to realize wireless root cause localization for poor voice quality by correlating UEMR data.After practical application,the MOS detection accuracy rate and the recall rate of quality defects have reached the level of industry-leading,and the poor quality optimization scheme was output.After optimization,the proportion of poor quality time in the cell uplink declines,with significant improvementeffect.
作者
顾竞雄
郑屹峰
林云
陈维新
周梁月
史超云
GU Jing-xiong;ZHENG Yi-feng;LIN Chong-yun;CHEN Wei-xin;ZHOU Liang-yue;SHI Chao-yun(China Mobile Group Zhejiang Co.,Ltd.,Hangzhou 310051,China)
出处
《电信工程技术与标准化》
2022年第S01期16-23,共8页
Telecom Engineering Technics and Standardization
