MOS Scores Explained: The Number Behind Call Quality

What MOS stands for

MOS is Mean Opinion Score. It grew out of the CCITT's (now ITU-T's) subjective testing work, formalized as the P.800 recommendation in 1996, as a way to quantify how humans perceive voice quality. The original method was simple: a group of listeners rated audio samples against P.800's five categories, and the average was the MOS.

• 5: Excellent. Like listening in the same room.
• 4: Good. Minor imperfections, but natural sounding.
• 3: Fair. Noticeable distortion or effort required to understand.
• 2: Poor. Difficult to understand.
• 1: Bad. Unusable.

Getting a panel of listeners together every time you want to test a call is obviously impractical. So the industry developed algorithmic models that predict what a human panel would score.

How modern MOS is calculated

Today, MOS is almost always computed by an algorithm rather than by human listeners. The two main families of approaches are:

PESQ (Perceptual Evaluation of Speech Quality, ITU-T P.862) compares the original audio signal to what came out the other end. It literally measures how much the audio was degraded. This is called an "intrusive" or "full-reference" method because it needs the original signal for comparison. PESQ is accurate but requires access to both sides of the conversation, which makes it impractical for live monitoring.

POLQA (Perceptual Objective Listening Quality Analysis, ITU-T P.863) is the successor to PESQ, designed for HD voice and modern codecs. Same basic idea: compare input to output. Higher accuracy with wideband and super-wideband audio.

E-model (ITU-T G.107) takes a different approach. Instead of analyzing the audio itself, it is a transmission-planning model that calculates expected conversational quality from network and codec impairments: one-way delay, packet loss, and an equipment impairment factor for the codec. Jitter is not a direct E-model input: it enters through the jitter buffer, as extra delay and as late packets discarded past the playout deadline. The E-model outputs an "R-factor" that theoretically ranges from 0 to 100, though the narrowband practical maximum is around 93-94. The R-factor maps to an estimated MOS. This is the method most monitoring tools use because it works with just network metrics and does not need the original audio.

These outputs carry different labels for a reason. A P.800 panel score is MOS-LQS (subjective listening quality), PESQ and POLQA produce MOS-LQO (objective listening-quality predictions), and the E-model's mapped score is MOS-CQE (estimated conversational quality). They are related but not interchangeable: a POLQA score and an E-model estimate for the same call measure different things and should not be compared as if they were the same number.

What the numbers mean in practice

Nobody gets a MOS of 5 on a real VoIP call. The codec alone introduces some degradation, so even a perfect network path tops out around 4.4 with common codecs like G.711. Here is a more realistic interpretation:

• 4.0 to 4.4: Excellent. Users are satisfied. This is the target for business VoIP.
• 3.6 to 4.0: Good. Occasional minor artifacts. Most users will not complain.
• 3.1 to 3.6: Acceptable for some use cases. Customer-facing calls will generate complaints.
• 2.6 to 3.1: Poor. Communication is strained. Users will actively seek alternatives.
• Below 2.6: Not viable for business use.

The important thing is that MOS is not linear. The difference between 4.0 and 3.5 is much more noticeable than the difference between 4.4 and 4.0. Small drops in the middle of the scale represent significant degradation in the listener's experience.

The codec ceiling

Every codec has a maximum MOS it can achieve even under perfect conditions. This is because compression always removes some information from the audio signal.

G.711 (the traditional PSTN codec, PCM with mu-law/A-law companding): Up to 4.4. This is the baseline for toll quality. It uses 64 kbps of bandwidth per direction.

G.729 (compressed, low bandwidth): Up to 3.9. Popular on bandwidth-constrained links because it only uses about 8 kbps. The tradeoff is that quality ceiling.

Opus (modern, adaptive): Up to about 4.5 at higher bitrates, which is the ceiling of the estimated-MOS scale itself. Compare that number with the two above carefully, though: the G.711 and G.729 ceilings are narrowband figures, while Opus is normally run as wideband or fullband audio, and narrowband and wideband scores are separate scales rather than points on one ladder. Opus can adjust its bitrate and complexity on the fly, which makes it excellent for variable network conditions. Most browser-based calling platforms use Opus.

If your monitoring shows a MOS of 3.9 and you are using G.729, you are actually at the codec's theoretical best. Switching to G.711 or Opus would raise the ceiling. But if you see 3.9 on G.711, something in the network is pulling quality down.

Network factors that lower MOS

The E-model gives us a clear picture of how each network issue impacts the score:

Packet loss has the largest impact per percentage point, but how large depends on the codec and on whether it conceals the packets it loses. Worked through the E-model, 1% random loss costs G.711 with packet loss concealment less than 0.1 MOS, while the same 1% on G.711 without concealment costs closer to 0.5. At 3% random loss, G.711 with concealment still lands around 4.1, where a low-rate codec like G.729 is down near 3.6. Bursty loss (several packets in a row) is worse than random loss at the same percentage because the jitter buffer cannot conceal gaps that long.

Latency has a gradual effect. One-way mouth-to-ear delay under 150ms is generally fine: that is the point below which interactivity is essentially transparent. Between 150ms and 400ms, conversations start to feel awkward because of the delay in responses, and 400ms is the ceiling for general network planning. Above that, people start talking over each other. The MOS impact is moderate compared to packet loss, but the conversational impact is real.

Jitter primarily affects MOS through its interaction with the jitter buffer. High jitter that exceeds the buffer creates effective packet loss. The jitter buffer itself adds latency. So high jitter forces a tradeoff: increase the buffer (more latency) or accept more discarded packets (lower quality).

Why a single number is both useful and dangerous

MOS gives you a single number to track, trend, and alert on. That is its strength. You can set a threshold (say, 3.8) and get notified when quality drops below it; our guide to monitoring MOS proactively covers where to put those thresholds so the alert fires before users complain. You can compare quality across locations, providers, or time periods.

The danger is in treating it as the whole story. A MOS of 3.8 caused by high latency is a very different problem from a MOS of 3.8 caused by packet loss. The user experience differs, the troubleshooting differs, and the fix differs. MOS tells you something is wrong. It does not tell you what.

This is why good VoIP monitoring always presents MOS alongside the underlying metrics. The score is a summary. The details are in the latency, jitter, and packet loss numbers that feed into it.

How we use MOS at VoIP Test

When we test your connection with our VoIP quality test, we measure the raw network conditions (packet loss, jitter, latency) and run them through the E-model to produce a MOS estimate, with measured jitter first translated into the buffer delay and late-discard loss it would cause, since G.107 does not take jitter directly. But we also show you the breakdown so you can see exactly which factor is pulling your score down. A single number is a starting point. The underlying data is where the answers live.

For a detailed look at the three metrics that feed into MOS, see Latency, Jitter, and Packet Loss: The Details in our VoIP From the Ground Up series.