Unpacking the meaning behind MOS
I have a confession. When I did my first Amazon Connect implementation, I learned about the term MOS (Mean Opinion Score).
Truthfully, I only knew that it was a way to measure voice quality. I even built part of the business case to shift to the cloud around how Amazon Connect provided best-in-class MOS. At the time, that’s all I needed to know. I didn’t need to know the technical behind it, but here at Operata, it’s one of the things we measure with our CX Observability platform, and so my curiosity about how it is measured drove me to get into much greater depth of MOS.
Forewarning, I get a little technical in this one but if you are in contact center ops and IT then I think it is worth getting your head around (like I probably should have). At the very least, it will help you with your BS detector when someone tries to dazzle you with jargon.
When someone asks you to rate voice quality on a scale of 1 to 5, they are really asking for the Mean Opinion Score, or MOS. That somewhat familiar question hides a world of behind-the-scenes data, audio processing, and network hiccups. This has nothing to do with quality monitoring of a contact center agent's interaction, which is based on process, compliance and tone. MOS is purely a technical measure of the clarity of the voice data.
In this article, I will take you inside the ITU-T E-Model behind modern MOS measurement (which is what Operata uses) to translate jitter, latency, and packet loss into a single, human-friendly MOS value. I’ll also explain the R-factor at the heart of the model, compare alternative MOS methods, and show you how observability uses MOS to inform your contact center and IT operations about voice quality.
Before I explain MOS I wanted to explain the R-factor which is an objective measure that takes into account many technical aspects, including signal-to-noise ratio, latency, and packet loss. The E-Model (ITU-T G.107) turns what the network is doing into a single quality index called R.
It is used within the MOS calculation of voice quality, quanitfying call quality on a scale from 0 to 100, with higher values indicating better quality
How R-Factor is Calculated:
R = Ro - Is - Id - Ie + A
Where:
MOS or Mean Opinion Score, which is a subjective measure of how good or bad a voice sounds under different data network conditions.
Early researchers recruited panels of listeners to rate call quality on a 1 to 5 scale ( 1 unusable and 5 perfect) as the data network used to carry voice was gradually made worse.
Subjective listening tests remain the ground truth, but they are slow and expensive. You cannot scale up hundreds of participants every time you want to spot network issues. That is why we calculate an estimated MOS in real time by using network metrics recorded on the agent’s computer. We approximate what a human listener would report. This automated MOS gives you continuous observability without interrupting callers or agents.
MOS uses three key data network performance metrics to derive a MOS value:
MOS Score:
To put this into context here are some call samples of what this sounds like:
While the E-Model is the go-to for live monitoring, several other standards exist to predict or measure voice quality:
Subjective Listening Tests (ITU-T P.800)
These are controlled lab experiments where human listeners rate call quality on a 1–5 scale. They remain the gold standard for accuracy but are slow, costly, and do not scale to production environments.
Full-Reference Objective Algorithms
These require both the original clean signal and the degraded signal to compute a perceptual difference. Two leading ITU-T standards are:
No-Reference Objective Algorithms
These require only the degraded signal without a clean reference.
Non-Intrusive Network Models (E-Model G.107 and extensions)
I learned that Operata uses an E-Model for a few good reasons. Firstly, its wideband/fullband extension estimates MOS solely from packet metrics. This is great because you get real-time, continuous quality scores without impacting live calls, making it perfect for CX Observability and your contact center.
In most contact center observability scenarios, the E-Model strikes the best balance of accuracy, scalability, and simplicity. You can, however, schedule periodic PESQ or POLQA tests for deeper validation in more controlled testing scenarios.
Our CX observability platform tracks metrics, including MOS, that can impact voice quality such as CPU, memory and network throughput.
These metrics tell you how your systems perform, but not how your callers experience voice quality. Operata CX observability bridges the gap.
Pair MOS with call volume, agent status, and trunk utilization to get a 360-degree view. You’re unlikely to ever miss a silent quality degradation again.
I warned you I was getting technical on this one but for good reason. Understanding the performance of your data network, for every second of each call is the foundation in understanding how good your customer and agent audio will sound.
MOS is the voice quality KPI that both contact center and IT can rally around and the E-Model’s R factor turns packet data into a number everyone understands.
You cna use MOS to prioritize incidents, verify fixes, and trend improvements. When you need gold-standard checks, run PESQ or POLQA.
Ready to start speaking MOS? Try Operata’s free trial today. See MOS live, set intelligent alerts, and solve voice quality issues before your callers notice.
Until next time and always,
Hooroo
Luke Jamieson brings over twenty years’ experience in contact centre leadership, having overseen some of Australia’s most successful operations and guided major CCaaS cloud migrations for leading organisations. Despite his track record, he remains committed to learning and is keen to share his discoveries each week at Operata. See more discoveries here.
Sweet sounds of voice innovation
Subscribe to our newsletter
You can unsubscribe at anytime. Read our privacy policy
Lock in some time with a Product Specialist.
