針對 Skype/SILK 網路語音封包遺失造成影響之使用者經驗導向的模型建立

Abstract

如何滿足使用者一直是應用軟體服務商必須面對的課題，而對於任何的服務商而言都必須了解使用者對於所提供之服務的滿意程度，以避免使用者因為不滿意的使用經驗而拒絕再次使用服務。在網路語音的範疇內，Skype可以說是目前最受歡迎的網路語音服務，同理應當瞭解其顧客對於Skype提供的網路語音品質的滿意程度。在可影響通話品質的因素之中，封包遺失無疑地對於即時網路語音的感受品質而言有著重大的影響，然而，針對封包遺失的感受品質影響卻甚少被討論，相關的使用者經驗模型也尚未被提出。已存在並廣泛運用的客觀感受模型POLQA，因其設計的限制並不適用於即時資料的傳輸，也因此若Skype想要預測其使用者對於服務的滿意程度，必須建立替代的使用者經驗模型才能達到目標。在此次研究中，我們模擬多種封包遺失狀況的Skype語音，並且在crowdsourcing平台請超過3000位受測者對各個模擬的Skype語音品質評分，經過分析可測量之網路品質指標（包含位元率、封包遺失率以及平均的單次封包遺失個數）與使用者經驗的關係，我們建立了一個以使用者經驗導向的模型，其適配程度(goodness of fit)為94.28%，在預測其他2獨立資料組分別得到92.16%和89%的適配程度，且這個模型與POLQA的預測比較之下，我們提出的使用者模型預測較POLQA更為精準。

Satisfying users has been an important issue for all application service providers, and all service providers shall understand how their users react to the services they provide. For real-time speech delivery services such as the most popular voice over IP (VoIP) service, Skype, packet loss is one of the major impairment to user satisfaction. However, there still lacks investigation and model of how users perceive the quality of speech with different packet loss conditions for Skype calls. A well-known perceptual objective model POLQA (Perceptual Objective Listening Quality Assessment) has its limitation to measure user experience for real-time data transmission, and hence, an alternative parametric model should be constructed for VoIP calls. Motivated by the existing model deficiency and lack of research concerning packet loss conditions, this study conducts a large-scale subjective experiment with more than 3000 participants involved on crowdsourcing platform to rate the perceived quality under different packet loss schemes. In the end, we propose a quality of experience (QoE) model with respect to measurable quality of service (QoS) parameters that are the indicators of network system performance including bitrate, loss rate and average burst length. The result shows that the goodness of fit of our proposed model is 94.28%, and we test the goodness of fit with 2 more independent datasets that obtain the goodness of fit being 92.16% and 89%, respectively. As our proposed model is compared to the POLQA, the proposed model outperforms POLQA modeling the perceived quality of Skype calls.