From Network Metrics to MOS: How Call-Quality Scores Work
Sooner or later a dashboard, a phone-system report, or a vendor datasheet will show you a number like MOS 3.2. It will be colored an accusing shade of orange. You will be expected to know what it means. MOS — the Mean Opinion Score — is the one number the entire voice world uses to summarize call quality on a scale of 1 to 5. It looks like a measurement. It behaves like a measurement. But it began life as something much stranger: a room full of humans listening to recordings and rating them, like wine judges for audio.
Understanding where the number comes from is what makes it useful. By the end of this guide you will know what the bands of the scale actually feel like on a real call. You will know how monitoring tools estimate a score from latency, loss, and jitter without any human listening. You will know why the codec changes the arithmetic and what a computed MOS fundamentally cannot know. And you will know how to use the score well: as a trend line and a translation device, never as a diagnosis.
This guide leans on the three underlying metrics without re-teaching them. If jitter is still a fuzzy concept, start with jitter explained and come back; MOS will make far more sense with that foundation in place.
In this guide
- Where the 1-to-5 scale comes from
- What the bands feel like on a real call
- How tools estimate MOS: the deduction model
- A worked example: watching a score degrade
- Why the codec moves the score
- What a computed MOS cannot hear
- Using MOS well: trend, alert, translate
- Where to go next
Where the 1-to-5 scale comes from
The original method was exactly as manual as it sounds. Telephone engineers played recorded speech samples — carried over whatever system was being tested — to panels of listeners. Each listener rated what they heard on a five-point scale: 5 excellent, 4 good, 3 fair, 2 poor, 1 bad. Average the opinions and you get the mean opinion score. The name is literal: it is the average of what people thought. Done with enough listeners under controlled conditions, it is still the gold standard. Call quality is ultimately a human experience and humans are the only authoritative instrument for it.
Panels of professional listeners do not scale to your Tuesday, so the industry built the next best thing. That means formulas that predict what the panel would have said, based on measurable network and codec facts. When your monitoring tool shows MOS 3.2, no one listened to anything. The tool measured latency, loss, and jitter and fed them into a standardized estimation model. Then it printed the score the model predicts a human panel would give. That distinction — measured opinion versus predicted opinion — is the key to everything else in this guide. That includes the moments when the prediction is confidently wrong.
You will meet these predicted scores in three places. Your phone system's admin portal typically records per-call quality statistics. Network monitoring tools estimate a score from their own test streams. And sometimes a voice provider's service commitments may promise to keep the score above a stated floor. All three are computed and all three rest on the same style of model — and knowing that lets you ask each one the right skeptical questions.
One quirk to absorb early: on the classic scale, the best achievable score is about 4.4, not 5.0. Ordinary phone-quality audio has a ceiling; even a flawless network carrying a standard call cannot reach a perfect score, because the encoding itself costs something and listeners can tell. If you see 4.3–4.4, you are looking at a call that cannot meaningfully be improved. Chasing 5.0 on a standard codec is chasing a number the math cannot produce.
What the bands feel like on a real call
Numbers only help if you can hear them in your head. Here is the scale translated into what the two people on the call actually experience:
| MOS band | How the call feels | What users do |
|---|---|---|
| 4.3 – 4.4 | Excellent. Clean, effortless, indistinguishable from the best landline call; you forget the technology exists. | Nothing. Nobody praises a phone call; silence is the trophy. |
| 4.0 – 4.3 | Good. A normal business call. A very picky listener might notice it is "a phone." | No complaints, no tickets. |
| 3.6 – 4.0 | Fair. Occasional artifacts — a clipped syllable, a brief flat robotic patch. Listening takes slight effort by the end of a long call. | The odd grumble in passing. Few formal tickets yet. |
| 3.1 – 3.6 | Poor territory. "Say that again?" becomes part of the conversation; people talk over each other when delay is the driver; robotic stretches when jitter is. | Complaints begin in earnest below about 3.5 and arrive as tickets. |
| Below 3.1 | Bad. Words go missing, meaning gets lost, the call is work. | Users hang up and redial hoping for better luck, switch to mobile phones, or move the meeting to chat. |
Two practical readings of that table. First, the useful range is narrow: the entire drama of call quality plays out between roughly 3.0 and 4.4. So a drop of half a point is a big event, not a rounding error. Second, the complaint threshold sits around 3.5 — but silence above it does not mean joy. A site living at 3.7 is paying a small tax on every conversation; nobody files a ticket, and everybody quietly hates the phones.
How tools estimate MOS: the deduction model
The standard estimation approach — the best-known version is called the E-model — works like a strict examiner grading a test. The call starts with a perfect score, and every impairment subtracts points. What is left at the end maps onto the 1-to-5 scale. (Many tools also show the intermediate grade, a 0–100 number called the R factor; a high-90s R corresponds to that 4.4 ceiling. Same idea, different ruler.) You do not need the equations; you need to know what subtracts, and roughly why.
Deduction one: the codec itself. Compressing speech costs fidelity before the network touches a single packet. A classic uncompressed phone codec starts near the ceiling; a high-compression codec starts several tenths lower. This is a flat, unavoidable entry fee, covered in more detail below.
Deduction two: delay. Latency does not distort the sound — it ruins the conversation. Below roughly 150 ms one-way, humans barely notice. Beyond that, the pause before each reply grows until both parties start talking at once, apologizing, and colliding again. The model subtracts gently at first, then steeply as the delay climbs toward the point where natural turn-taking breaks down. Note that this is total mouth-to-ear delay: network latency plus the jitter buffer's holding time.
Deduction three: loss. Every lost packet is a missing slice of speech. Codecs conceal small gaps impressively — a single vanished 20 ms slice is patched so well most listeners never notice. But concealment fails as loss rises. It fails faster when losses arrive in bursts, because a run of missing slices defeats any patching trick. The deduction therefore depends on how much loss, how bursty, and how fragile the codec is.
Deduction four: jitter — by conversion. Here is the subtle one. Jitter never appears in the formula directly. By the time audio reaches an ear, the jitter buffer has converted it into the two currencies above. The buffer holds packets to smooth the rhythm, which adds delay. Any packet arriving too late for its moment is discarded, which adds loss. A tool estimating MOS models that conversion — assuming a buffer of some size and policy — and charges the delay and loss accounts accordingly. This is why two tools watching the same stream can disagree slightly: they may assume different buffers.
That is the whole machine: start from the ceiling, pay the codec's entry fee, subtract for delay, subtract for real loss plus the loss jitter manufactured. And map the remainder onto the 1-to-5 scale. In reality the deductions interact a little. Impairments gang up, and the same loss hurts more on a call already strained by delay. But the additive ledger is accurate enough to reason with, and it is exactly how the tools present their verdicts. Every MOS estimate you will ever read is some version of it.
A worked example: watching a score degrade
Suppose your measurements for a branch office say this: one-way delay is around 120 ms once the jitter buffer's contribution is included. Packet loss is 1%. Jitter averages 20 ms with spikes to 40 ms. In this example, calls use a high-compression codec to save WAN bandwidth. Here is the examiner's ledger, in deliberately rounded, illustrative numbers — real implementations differ in the decimals but not in the shape:
Starting point (clean network, uncompressed codec) 4.4
Codec entry fee: high-compression codec -0.3 => 4.1
Delay: ~120 ms one-way, buffer included -0.1 => 4.0
Jitter: 20-40 ms vs. a ~40 ms buffer
=> late discards; effective loss 1% -> ~2% -- (feeds next line)
Loss: ~2% effective, on a loss-sensitive codec -0.5 => 3.5
Estimated MOS 3.5 (complaints begin here)
Walk the steps and each is intuitive. The codec spends 0.3 before the first packet flies. The delay is comfortably under the 150 ms line, so conversation flows and the charge is small. The jitter mostly fits inside the buffer, but the 40 ms spikes push occasional packets past their deadline, roughly doubling the loss the listener effectively experiences. And two percent effective loss on a fragile codec is a real wound: half a point. Result: 3.5, a call that works but grates — and a branch that is one bad backup window away from a ticket storm.
The ledger also shows you the repair options in priority order, free of charge. Cutting the jitter spikes (or letting the buffer breathe) claws back most of that 0.5. Switching the codec buys 0.3 at the cost of bandwidth. Shaving delay buys little, because delay was never the problem here. This is the quiet superpower of understanding the model: a MOS number tells you how bad. The ledger tells you why, and what to fix first.
Why the codec moves the score
A codec is the algorithm that turns sound into packets and back. Codecs differ along one great trade-off: how much bandwidth they use versus how much fidelity they keep and how gracefully they suffer. For MOS purposes, three general characteristics matter.
Compression lowers the starting score. A high-compression codec squeezes speech into a fraction of the bits, which is a gift to thin WAN links. The squeezing discards audio information listeners can hear missing. The codec starts the exam with points already gone.
Compression raises the cost of every lost packet. In a lightly compressed stream, each packet carries relatively redundant audio, and concealment patches a gap convincingly. In a heavily compressed stream, every packet is dense, irreplaceable information. And worse, decoders often need context from previous packets, so one loss can smear damage across the packets that follow. The same 1% loss therefore charges the compressed call more. Loss-sensitivity and starting score travel together: the codecs that start lower also fall faster.
Wideband raises the ceiling. Classic telephone audio deliberately carries only a narrow slice of the frequencies in a human voice — the familiar "phone sound." Wideband codecs carry a broader slice, and speech through them sounds noticeably more present and natural. Scoring scales acknowledge this by extending the ceiling: a clean wideband call can score where narrowband cannot reach. The practical note for your dashboards: compare like with like. A wideband call at 4.1 and a narrowband call at 4.1 did not sound the same. A codec change can move your MOS trend without the network changing at all.
What a computed MOS cannot hear
A monitoring tool's MOS is inferred entirely from network arithmetic — the loss, delay, and jitter of packets observed at some measurement point. That makes it tireless and consistent, and blind in specific, predictable ways. It cannot hear acoustic echo from a bad speakerphone. It cannot hear a five-dollar microphone, a headset with a dying battery, aggressive noise suppression eating quiet speakers, or the boss taking the call on speaker in a stairwell. It cannot see the CPU-starved laptop that received every packet perfectly and then glitched the audio while decoding it. All of that arrives at the ear as "a bad call" and none of it appears in packet statistics. So users can be miserable while the dashboard glows green, and both are telling the truth.
The disagreement runs the other way too. The dashboard can show an ugly red dip while nobody complains, because the measured path is not exactly the path calls took. For example, a synthetic probe to the data center suffered while actual calls flowed peer-to-peer around the trouble. Or the impairment struck during silence, or a lunch hour, or fell on the tolerant ears of people who simply redialed. A computed MOS describes network conditions expressed in call-quality units. That is genuinely valuable; it is just not the same thing as an ear.
It also matters which computed MOS you are reading, because there are two flavors. A synthetic score comes from a test stream between measurement points. It watches the path around the clock, even at 3 AM with no calls to observe. But it grades a stand-in stream rather than any real conversation. An endpoint-reported score comes from the phones and apps themselves. They count the loss, jitter, and delay they actually experienced during real calls and report it back. That is ground truth for the media path, but it exists only when calls happen. The endpoint-reported score is just as deaf to microphones and stairwells as every other formula. The two are teammates: synthetic scores stand guard between calls, and endpoint reports confirm what real calls lived through. And a gap between them is itself a clue that the test path and the call path differ.
Tip: MOS is per direction. Each side of a call gets its own score, and impairments are frequently one-sided — an upload-choked link ruins outbound audio while inbound stays pristine. The complaint itself carries the clue: "you sounded like a robot" convicts your send direction. When a user and the dashboard disagree, first check whether you are looking at the direction they were listening to.
Using MOS well: trend, alert, translate
Used naively — one number, one moment, one threshold — MOS mostly generates arguments. Used well, it does three jobs better than any other metric in the building.
Trend it per site and per link. A single MOS reading is weak evidence; a line of them is a story. Chart the score per branch, per WAN link, per week. Healthy sites draw a boring flat line in the low 4s. Treasure that boring line, because the day it sags to 3.8 you know something changed even though nobody has complained yet. The trend turns MOS from a verdict into an early-warning instrument.
Alert on sustained drops, not single dips. One rough sample during one backup burst is weather. Fifteen minutes below your floor — say 3.8, chosen to sit above the 3.5 complaint line — is climate, and worth an alert. Alerting on every momentary dip teaches everyone to ignore the alerts, which is worse than not having them.
Correlate dips with the metrics that drive them. When the score sags, the very next click should be the underlying three: did jitter rise, did loss appear, did delay step up? Line the MOS timeline against those measurements — gathered as described in measuring jitter properly — and against link utilization for the same minutes. The score tells you when it got bad; the metrics tell you why; and the thresholds in jitter targets by workload tell you which metric crossed its own line first.
Translate for management. Executives do not feel milliseconds; everyone understands "call quality dropped from 4.2 to 3.4 at the branch every morning at 9:40." A MOS trend chart is the rare network graph that needs no interpreter. That makes it the natural language for status reports and, when the time comes, for budget requests. "This line stays above 4.0 after the upgrade" is a promise a non-engineer can hold you to.
Where to go next
Keep MOS in its proper job. It is a summary statistic — latency, loss, and jitter compressed into one human-shaped number, resting on assumptions about codecs and buffers. Summaries are for noticing and for communicating: trend it, alert on sustained sags, and hand the chart to management. Diagnosis belongs to the three underlying metrics. Repair belongs to the practices covered across the rest of the Jitter & Real-Time Traffic guides. Those practices include measuring jitter credibly and holding it to defensible targets. They include understanding the buffers that trade delay for smoothness, and protecting the voice lane with a starter QoS design. Report with MOS; troubleshoot with the metrics. Teams that keep those two directions straight fix calls faster and argue less.
Put this into practice with SysPeep Monitor
A sagging call score is a symptom; the cause lives in the metrics underneath. SysPeep Monitor tracks the link utilization, errors, discards, and device health that drive call quality — over SNMP, WMI, and SSH. Its dashboards and alert thresholds flag the sustained change, not every blip. The free trial installs on a Windows machine in minutes.
Download Free TrialLearn About SysPeep