基于5G语音质差自适应算法研究及应用

doi:10.11959/j.issn.1000-0801.2023249

摘要/Abstract

摘要：

MOS 通常被业界用于评价语音质量，它能够客观公正地反映用户语音业务的感知。通过路测获取数据的方式难度大、成本高，通常采用训练好的监督学习模型预测MOS。但运营商语音数据存在MOS低分数据占比低和时序变化的特性，这种数据特性影响了模型预测的精度和泛化性。在研究现有运营商数据采集系统和机器学习算法的基础上，提出了一种面向5G语音质差MOS评估的自适应算法。首先，基于全参评估的 POLQA 算法测试设备获取训练数据，保证了训练样本的准确性；其次，通过数据增强的方法，解决了质差样本获取难度大的问题；最后，基于自适应算法选型实现周期性动态地根据数据特征的时序变化选择最佳MOS预测模型，实现5G语音质量规模化、智能化的评估。

关键词: 5G语音质量, MOS, 机器学习, 自适应

Abstract:

MOS (mean opinion score) is usually used to evaluate voice quality in the industry.It can objectively and fairly reflect the user’s voice service perception.It is difficult and costly to obtain data by road test, so a trained supervised learning model is usually used to predict the MOS score.However, the operator voice data has the characteristics of low percentage of MOS low score data and time sequence change, which affects the accuracy and generalization of the model prediction.Based on the study of existing data acquisition systems and machine learning algorithms of operators, an adaptive algorithm for MOS evaluation of 5G speech quality was proposed.Firstly, POLQA algorithm test equipment based on full parameter evaluation obtained training data to ensure the accuracy of training samples.Secondly, by means of data enhancement, the difficulty of acquiring poor quality samples was solved.Finally, based on the adaptive algorithm selection, the optimal MOS prediction model could be selected periodically and dynamically according to the timing changes of data features, so as to achieve large-scale and intelligent evaluation of 5G voice quality.

Key words: 5G voice quality, MOS, machine learning, adaptive

中图分类号:

TN915

赵宇翔, 纪雅欣, 余立, 周天一, 周航. 基于5G语音质差自适应算法研究及应用[J]. 电信科学, 2023, 39(11): 153-163.

Yuxiang ZHAO, Yaxin JI, Li YU, Tianyi ZHOU, Hang ZHOU. Research and application of adaptive algorithm for 5G voice quality evaluation[J]. Telecommunications Science, 2023, 39(11): 153-163.

图/表 15

图1

表1

图2

图3

图4

图5

图6

图7

图8

表2

表3

图9

表4

表5

表6

参考文献 29

[1]	STREIJL R C , WINKLER S , HANDS D S . Mean opinion score (MOS) revisited:methods and applications,limitations and alternatives[J]. Multimedia Systems, 2016,22(2): 213-227.
[2]	玉荣娟 . 语音质量评估及其优化策略[J]. 移动通信, 2009,33(20): 79-81.
	YU R J . Speech quality assessment and optimization strategies[J]. Mobile Communications, 2009,33(20): 79-81.
[3]	陈明 . VoLTE 语音质量评估的 POLQA 算法研究[J]. 移动通信, 2016,40(15): 20-25.
	CHEN M . Research on the POLQA algorithm for VoLTE speech quality assessment[J]. Mobile Communications, 2016,40(15): 20-25.
[4]	薛晓宇, 龙杰, 方义成 ,等. 5G VoNR语音关键技术分析[J]. 电信工程技术与标准化, 2022,35(9): 81-85.
	XUE X Y , LONG J , FANG Y C ,et al. Analysis of key technologies for 5G VoNR speech[J]. Telecommunications Engineering Technology and Standardization, 2022,35(9): 81-85.
[5]	吕军 . 中国电信FDD 4G VoLTE语音感知MOS质量问题分析和优化提升的研究[J]. 数据通信, 2020(5): 1-9.
	LYU J . Analysis and optimization of MOS quality problems in China Telecom FDD 4G VoLTE voice perception[J]. Data Communication, 2020(5): 1-9.
[6]	张磊, 张恒, 李一 ,等. CSFB 语音业务质量评估与优化方法研究[J]. 邮电设计技术, 2016(2): 66-72.
	ZHANG L , ZHANG H , LI Y ,et al. Research on the quality evaluation and optimization method of CSFB voice service[J]. Posts and Telecommunications Design Technology, 2016(2): 66-72.
[7]	李荪, 曹峰, 刘姿杉 . 面向算法模型的语音数据集质量评估方法研究[J]. 计算机科学, 2022,49(S2): 519-524.
	LI S , CAO F , LIU Z S . Research on the quality evaluation method of speech datasets for algorithm models[J]. Computer Science, 2022,49(S2): 519-524.
[8]	辛建国, 刘洪波, 王浩 . VoLTE语音质量（MOS）提升研究[J]. 山东通信技术, 2021,41(1): 40-45.
	XIN J G , LIU H B , WANG H . Research on improving VoLTE voice quality (MOS)[J]. Shandong Communication Technology, 2021,41(1): 40-45.
[9]	ZEZARIO R E , FU S W , CHEN F ,et al. Deep learning-based non-intrusive multi-objective speech assessment model with cross-domain features[J]. IEEE/ACM Transactions on Audio, 2023,31(1): 54-70.
[10]	FU S W , TSAO Y , HWANG H T . Quality-net:an end to-end non-intrusive speech quality assessment model based on BLSTM[J]. Inter Speech, 2018(1): 1873-1877.
[11]	李伦彬, 滕海坤, 王诗莹 . 一种用于语音质量评价的半监督学习方法研究[J]. 长春工程学院学报(自然科学版), 2021,22(3): 119-124.
	LI L B , TENG H K , WANG S Y . Research on a semi-supervised learning method for speech quality evaluation[J]. Journal of Changchun Institute of Technology (Natural Science Edition), 2021,22(3): 119-124.
[12]	陈森, 金岩华 . VoLTE 端到端语音质量评估及优化研究[J]. 无线互联科技, 2018,15(3): 10-11,20.
	CHEN S , JIN Y H . Research on VoLTE end-to-end voice quality evaluation and optimization[J]. Wireless Internet Technology, 2018,15(3): 10-11,20.
[13]	王希 . 基于 MR 数据与机器学习的 LTE 用户感知评估方法[J]. 移动通信, 2018,42(8): 21-26.
	WANG X . LTE user perception evaluation method based on MR data and machine learning[J]. Mobile Communications, 2018,42(8): 21-26.
[14]	黄鹂声, 冉金也, 罗静 ,等. 基于XDR数据分析的OTT视频服务感知质量评估方法[J]. 计算机研究与发展, 2021,58(2): 418-426.
	HUANG L S , RAN J Y , LUO J ,et al. A method for evaluating the perceived quality of OTT video services based on XDR data analysis[J]. Computer Research and Development, 2021,58(2): 418-426.
[15]	张瑜, 吴凯, 郭杰 ,等. 基于数据质量评估的自适应序贯航迹关联算法[J]. 系统工程与电子技术, 2022,44(11): 3477-3485.
	ZHANG Y , WU K , GUO J ,et al. Adaptive sequential track association algorithm based on data quality assessment[J]. Systems Engineering and Electronic Technology, 2022,44(11): 3477-3485.
[16]	孙艳丰, 陈亮, 胡永利 . 基于对比学习的双分类器无监督域适配模型[J]. 北京工业大学学报, 2023,49(2): 197-204.
	SUN Y F , CHEN L , HU Y L . The unsupervised domain adaptation model of double classifiers based on contrastive learning[J]. Journal of Beijing University of Technology, 2023,49(2): 197-204.
[17]	TAO X M , LI Q , GUO W J ,et al. Adaptive weighted over-sampling for imbalanced datasets based on density peaks clustering with heuristic filtering[J]. Information Sciences, 2020(519): 43-73.
[18]	张鸿, 黄保坤 . 拉曼光谱神经网络分类与样本扩充方法[J]. 福建电脑, 2023,39(4): 19-24.
	ZHANG H , HUANG B K . Neural network classification and sample expansion in Raman spectrum[J]. Fujian Computer, 2023,39(4): 19-24.
[19]	李娜, 李琳琳, 王永远 . 麦克风自适应算法在鲁棒语音合成中的应用研究[J]. 中国科技信息, 2013(11): 127,130.
	LI N , LI L L , WANG Y Y . Research on the application of microphone adaptive algorithm in robust speech synthesis[J]. China Science and Technology Information, 2013(11): 127,130.
[20]	朱永宽, 谷涓涓 . 自适应遗传算法在聚类分析中的应用[J]. 黑龙江科技信息, 2010(25): 52-53.
	ZHU Y K , GU J J . Application of adaptive genetic algorithm in cluster analysis[J]. Heilongjiang Science and Technology Information, 2010(25): 52-53.
[21]	王超, 田波, 李子睿 ,等. 基于自注意力子域自适应对抗网络的轴承故障诊断方法[J]. 中国舰船研究, 2023(6): 1-10.
	WANG C , TIAN B , LI Z R ,et al. A bearing fault diagnosis method based on self attention subdomain adaptive adversarial network[J]. China Shipbuilding Research, 2023(6): 1-10.
[22]	谢顺文, 林高全, 邱宏洪 ,等. 一种5G上行覆盖增强的方法[J]. 移动通信, 2023,47(5): 90-96.
	XIE S W , LIN G Q , QIU H H ,et al. A method for enhancing 5G uplink coverage[J]. Mobile Communications, 2023,47(5): 90-96.
[23]	彭艺, 魏翔, 朱桢以 ,等. 基于能量检测与信道质量评估的动态信道接入算法[J]. 数据通信, 2023(1): 5-9.
	PENG Y , WEI X , ZHU Z Y ,et al. Dynamic channel access algorithm based on energy detection and channel quality evaluation[J]. Data Communication, 2023(1): 5-9.
[24]	ITU. ITU-T Recommendation:P.863,Perceptual objective listening quality assessment (POLQA)[S]. 2011.
[25]	PAN Z Q , YU W J , YI X K ,et al. Recent progress on generative adversarial network (GAN):a survey[J]. IEEE Access, 2019(7): 36322-36333.
[26]	KIM J Y , BU S J , CHO S B . Zero-day malware detection using transferred generative adversarial networks based on deep autoencoders[J]. Information Sciences, 2018(460/461): 83-102.
[27]	RADFORD A , METZ L , CHINTALA S . Unsupervised representation learning with deep convolutional generative adversarial networks[J]. CoRR, 2015(1): 1-10.
[28]	GOODFELLOW I , POUGET-ABADIE J , MIRZA M ,et al. Generative adversarial nets[J]. Mining of Massive Datasets, 2014(1): 1-10.
[29]	周文豪, 牛英滔, 施育鑫 . 无线通信中的机器学习算法综述[J]. 无线电通信技术, 2022,48(6): 1049-1057.
	ZHOU W H , NIU Y T , SHI Y X . Overview of machine learning algorithms in wireless communication[J]. Radio Communication Technology, 2022,48(6): 1049-1057.

业务类型	建议值
MOS良好区间	[3.8, 4.8]
MOS欠质差区间	[3.0, 3.8)
MOS质差区间	[1.08, 3.0)

业务特征	描述
信令面特征	接入和承载建立特征	5G承载建立标识
	5G信令特征	信令面失败次数
用户面特征	业务行为特征	5G业务时长
		5G业务占比
	业务质量KQI特征	上/下行RTP丢包数
		上/下行RTP期望包数
		本周期收到的最后一个RTP包相对于开始时间的时延
		切片周期内RTP连续丢包大于或等于2且小于6的次数
其他特征	标识信息	MSISDN码
		SESSION ID
		XDR ID
	Interface字段	Gm、S1-U、N3接口

序号	特征名称	含义
1	AvgInterDelay	平均相对时延
2	MaxIntervalNonsid	RTP最大包间隔
3	MaxInterDelay	最大相对时延
4	MaxJitter	最大RTP抖动
5	MaxLossRate	RTP最大连续丢包率
6	LossRateNonsid	RTP丢包率（非静默帧）
7	RtcpPacketLossRateKeeptime	RTCP单位时间的丢包数
8	MaxLoss	RTP最大连续丢包数
9	MaxLossNonsid	RTP最大连续丢包数（非静默帧）
10	CodeRateFeat	编码速率

指标	数据	总体	质差	欠质差	良好
召回率	原始数据	0.582	0.518	0.488	0.739
	增强后数据	0.772	0.793	0.779	0.745
精准度	原始数据	0.731	0.876	0.376	0.824
	增强后数据	0.788	0.978	0.533	0.852
F1值	原始数据	0.648	0.651	0.276	0.779
	增强后数据	0.768	0.876	0.472	0.795

模型	总体	质差	欠质差	良好
XGBoost	0.782	0.968	0.528	0.850
LightGBM	0.705	0.952	0.313	0.843
CatBoost	0.788	0.978	0.533	0.852
CNN	0.759	0.934	0.374	0.963
DNN	0.759	0.902	0.374	0.999
LSTM	0.673	0.842	0.227	0.852
Bi-LSTM+Attention	0.692	0.823	0.212	0.868
平均法	0.737	0.914	0.366	0.889
stacking₁	0.804	0.970	0.623	0.818
stacking₂	0.793	0.971	0.602	0.806
stacking₃	0.821	0.964	0.676	0.823