What Is Mean Opinion Score (MOS)? A 2026 Guide for Voice AI

Your dashboard reports a Mean Opinion Score of 4.2, and you still have no way to act on it. The number arrives with no scale identifier, no test conditions, and no sign that a human ever listened to anything. Here’s how MOS works and how to use it when testing your own AI voice agents.

TL;DR: Mean Opinion Score

• MOS is nine separate metrics sharing one acronym, and ITU-T P.800.1 gives each one its own identifier for listening, conversational, or talking quality.
• The score on most call dashboards is MOS-CQE from the G.107 E-model, a planning estimate that tops out at 4.5 and was never scoped to judge individual customer calls.
• PESQ was deleted from the ITU catalog on 5 January 2024, though plenty of tools still ship it under the same name.
• Published TTS naturalness scores are not comparable between vendors, because most papers omit the scale labels, increments, and listener instructions that move the number.
• The metric says nothing about whether your agent captured the callback number or completed the booking.

What is a Mean Opinion Score (MOS)?

Mean Opinion Score is the arithmetic mean of quality ratings that listeners assign to a speech sample on a five-point scale. ITU-T P.800 set the method in 1996 for telephone transmission quality, and everything called MOS today descends from it.

The original setup was a controlled listening room. Subjects heard short recorded samples passed through a system under test, then rated each one. Averaging those ratings gave a single number for the system.

The metric describes a system measured over many samples. One call sits well outside what the method was built to judge.

The five-point absolute category rating scale

The absolute category rating (ACR) scale maps five verbal labels onto five numbers.

Listeners pick a label. The score is the mean of those picks across everyone who rated that condition.

Variance disappears in that average, so a condition where half the panel said Excellent and half said Poor reads identically to one where everyone said Fair. Distance between labels also gets treated as equal, which is a documented limitation of the method.

What a bare MOS number leaves out

A MOS value with no context is close to unreadable. You cannot tell whether humans produced it or an algorithm predicted it, whether it describes listening or conversation, or what audio bandwidth was in play.

P.800.1 states that its own shorthand identifiers are insufficient to describe test context, and points to P.800.2 for full reporting.

The nine MOS types defined in ITU-T P.800.1

ITU-T P.800.1, in force since July 2016, splits the score across two axes. One axis is what got rated, and the other is how the score was produced.

There are nine identifiers and one acronym. A vendor reporting "MOS 4.3" has told you which of nine numbers it is only if the suffix came along with it.

Subjective, objective, and estimated scores

The three rows rank by how close each is to an actual human judgment.

• Subjective scores come from people rating audio under P.800 conditions.
• Objective scores come from perceptual models that analyze the waveform and predict what a panel would have said.
• Estimated scores come from planning models that never touch audio and work from network parameters alone.

Accuracy drops as you move down that list. An estimated score is a model of a model of an opinion.

Bandwidth suffixes and why they change the number

P.800.1 appends a bandwidth letter to each identifier. N covers narrowband audio from 300 to 3400 Hz. W covers wideband from 50 to 7000 Hz, and S and F cover super-wideband and fullband.

A listener rates against a reference at that bandwidth. The same clip scored as MOS-LQON and MOS-LQOW produces two different numbers, and neither one is wrong.

How a MOS listening test is run

Running a real MOS test means recruiting listeners, controlling their environment, and choosing a rating procedure. The procedure you pick determines what the resulting number can support.

Absolute, degradation, and comparison category rating

Three procedures dominate, and they answer different questions.

• Absolute category rating (ACR): listeners hear one sample and rate it 1 to 5 with no reference. This produces classic MOS.
• Degradation category rating (DCR): listeners hear a clean reference, then the processed version, and rate how noticeable the degradation was.
• Comparison category rating (CCR): listeners hear two samples in random order and rate one against the other.

ACR is the most common and the least sensitive. When two systems are close in quality, DCR and CCR separate them where ACR flattens them together.

Crowdsourced listening tests under ITU-T P.808

Lab panels are slow and expensive, so the ITU standardized the online version. ITU-T P.808 covers ACR, DCR, and CCR run through crowdsourcing platforms, with listener qualification, hearing screens, and environment checks written into the procedure.

P.808 sits alongside P.800 as a complement. The lab is better controlled, while the crowd covers a wider range of real devices and rooms, which can raise the external validity of the result.

Microsoft publishes an open-source P.808 implementation on Mechanical Turk, which is the fastest path to running your own listening test on agent audio.

Listener count and score stability

Small panels produce unstable means. Research feeding into P.808 measured MOS reliability against vote count, comparing crowdsourced results against a standard laboratory experiment across three speech datasets.

A MOS reported without its panel size cannot be weighed against anyone else's.

How software estimates MOS without human listeners

Nobody runs a listening panel on every production call. Three families of software fill the gap, and each one produces a different row of the P.800.1 grid.

Full-reference models POLQA and PESQ

Full-reference models compare a degraded signal against the clean original. ITU-T P.863, known as POLQA, is the current recommendation and outputs MOS-LQO in narrowband and super-wideband modes.

PESQ was superseded years ago, and the ITU deleted the entire P.862 family from its catalog on 5 January 2024, pointing users to P.863 in its place.

Tools still ship PESQ, because it is free while POLQA is licensed. A PESQ number in a 2026 datasheet is a score from a withdrawn standard, and the version and implementation both need stating before the figure means anything.

Full-reference models carry a structural limit for voice AI. They need the clean original, and a live customer call has no clean original to compare against.

Non-intrusive models: P.563, DNSMOS, and NISQA

Non-intrusive models score the degraded audio on their own. ITU-T P.563 was the first standardized attempt, built for narrowband telephony, and it only covers narrowband signals.

Its accuracy drops further on conversational speech and modern VoIP distortions that did not exist when it was written.

Neural predictors solved most of that problem. DNSMOS P.835 scores speech quality, background noise, and overall quality as three separate outputs, trained on crowdsourced ratings collected under ITU-T P.835. NISQA predicts multiple quality dimensions from a CNN with self-attention.

These are what production monitoring runs, and they are also predictions of a panel that never sat down. Treat small movements as noise until you have the model's own reported error margin in front of you.

The E-model in ITU-T G.107 and MOS-CQE

The E-model in ITU-T G.107 produces most of the MOS numbers on call dashboards, and it never analyzes audio. It computes a transmission rating factor R from network and equipment parameters, then converts R into an estimated conversational score, MOS-CQE.

R is built by subtraction. Signal-to-noise ratio sets the starting point, and impairments for delay, echo, codec choice, and packet loss come off the top. With every parameter at default, R lands at 93.2.

The scale stops at 4.5. G.107 Annex B defines MOS-CQE as 4.5 for any R of 100 or above. A dashboard using the E-model cannot report a 5.0, and a 4.5 means the model ran out of room.

The model adds points for user tolerance. G.107 includes an advantage factor that credits a connection for being convenient.

Those points are added to R before the MOS conversion. A satellite call and a landline call with identical impairments will report different MOS values by design.

The ITU scoped it to planning. The G.107 scope clause states that R-to-opinion transforms exist for transmission planning purposes and are not for actual customer opinion prediction, and notes that no agreed-upon model for the latter exists.

A threshold that pages someone when one call's MOS drops below 3.5 is applying a network capacity planning tool to an individual customer conversation.

MOS score ranges and user satisfaction bands

The bands everyone quotes come from one table in G.107, and they describe R-value planning targets with MOS-CQE attached.

The familiar "3.5 is acceptable" heuristic sits between two rows that the ITU labels as some users dissatisfied and many users dissatisfied. G.107 presents the whole table as provisional planning guidance.

There are three consequences for anyone setting a threshold on agent audio:

• The bands assume 3.1 kHz handset telephony. A wideband WebRTC agent sits outside the conditions the table was built for.
• The percentages are population estimates. 73% good-or-better at R 70 describes a modeled user base, and never the caller currently on the line.
• Risk should set the number. A medical intake call and an order-status call have different tolerances, so one global cutoff serves neither well.

Set thresholds per workflow, then validate them against your own call outcomes. Our guide to voice quality testing covers how to build acceptance criteria per call flow.

Limits of MOS for AI voice agents

This metric was designed for a codec or a network carrying human speech. An AI voice agent generates novel audio every turn and runs a chain of models before that audio exists.

MOS coverage stops at the audio layer

A call can score 4.4 and still be a total loss. The agent may have transcribed a policy number wrong, skipped verification, or hung up before a transfer landed.

None of that touches the waveform. Audio quality and task success are separate axes, and MOS reads only one of them.

The inverse happens too. A call with mild jitter can complete a booking perfectly, and paging someone over its MOS wastes an engineer's afternoon.

Vendor naturalness scores are not comparable across labs

TTS vendors publish MOS figures that are close to incomparable. An SSW 2023 analysis of Interspeech and SSW papers from 2021 to 2022 found most authors omitted the scale labels, the increments, and the listener instructions.

Where those details did appear, implementations diverged. The same analysis showed how much the wording of the question matters: papers asking listeners about quality and papers asking about naturalness are not measuring the same thing, so two vendors can publish different numbers for one voice.

Treat a published TTS MOS as a claim about one lab's setup. Our rundown of TTS voice options covers what separates these systems in production.

A per-call MOS is a planning estimate

An E-model MOS responds to jitter, packet loss, and delay. It cannot respond to a robotic TTS render, a mispronounced drug name, or an agent talking over a caller. Those problems live in the AI pipeline, where a network planning model has no input.

Metrics to pair with MOS for voice agents

Each MOS type covers a specific slice of the problem, and every gap needs its own signal.

Audio metrics answer whether the call sounded acceptable, and conversation metrics answer whether it worked. Cekura's evaluation metrics guide covers the second column in depth, including speaking rate, talk ratio, and interruption recovery.

Five practices for using MOS in voice AI testing

1. Record the full identifier and the test conditions

Log the P.800.1 identifier, the bandwidth suffix, the model or panel that produced the score, and the codec in use. A stored value of "4.2" with no metadata is unusable six months later when you need to know whether quality regressed.

2. Compare inside one harness and never across vendors

MOS is only meaningful relative to other scores from the same setup. Two vendors running different models, different bandwidths, and different reference conditions produce numbers you cannot subtract from each other.

We built the Cekura voice agent benchmark on exactly this rule. Seven platform configurations run 82 matched scenarios three times each under one fixed harness, each provider choosing its own models, speech components, and settings, with published methodology and shareable run reports.

3. Set thresholds per workflow risk level

A single global cutoff pages too often on low-stakes calls and too late on regulated ones. Anchor each threshold to what goes wrong when the workflow degrades.

Payment collection, clinical intake, and identity verification earn tighter bands. Store hours and order status can run looser.

4. Alert on sustained patterns across several samples

One degraded call is noise, but a pattern across consecutive samples is a regression worth your team’s time.

Requiring three consecutive samples below threshold before an alert fires is what keeps an alerting system from burning its credibility in the first week.

5. Pair every audio score with a task outcome

Score the audio and the outcome together, then look at the calls where the two disagree. High MOS with a missed booking points at prompt or tool logic, and low MOS with a completed task tells you the threshold is too tight.

That disagreement set is the most useful queue in voice QA, and it exists only when both numbers sit on the same call record.

How Cekura measures voice quality for AI agents

Cekura scores voice quality from the audio channel directly, alongside the conversation-level signals MOS cannot reach.

The public Cekura benchmark shows the approach in the open. Voice Tone + Clarity ran from ElevenLabs at 4.47 out of 5 down to Pipecat at 3.74, with Retell and LiveKit tied at 4.36.

Across the six configurations scored on voice tone, perceived quality was within 0.73 points end to end, while three-run repeatability across those same six ran from Retell at 75.61% down to Vapi at 59.76%.

The seventh configuration, Gemini Live, sat at 30.49% and is excluded from the voice tone comparison.

Sounding good and working reliably move independently, which is why you need to measure both.

Pre-production

• Scenario simulations across booking, rescheduling, refunds, and identity verification run before a change reaches callers.
• Regression suites triggered on prompt, model, or provider changes through CI.

Infrastructure

• Interruption, background noise, packet-level degradation, and voice activity detection behavior tested under controlled conditions.
• Latency measured per component, so a slow turn traces to STT, the model, or TTS.

Observability

• Production call scoring against custom metrics you define in plain language, with dropout points and escalation patterns surfaced in dashboards.

Native integrations work out of the box for Retell, Vapi, ElevenLabs, LiveKit, Pipecat, Bland, and more. You add a testing and observability layer over the stack you already run.

Cekura supports SOC 2, HIPAA, and GDPR compliance, covering transcript redaction, role-based access, and audit trails.

For the wider monitoring layer this sits inside, see our voice observability guide.

Where Mean Opinion Score fits in your voice AI stack

Mean Opinion Score is a real measurement with a narrow job. It tells you whether audio sounded acceptable to an average listener under stated conditions, and it does that well when you report the identifier, the method, and the panel or model behind it.

It was never built to judge whether an AI agent understood a caller or finished a task. Those questions need conversation-level testing before launch and call-level scoring after it.

Which of your call flows would sound fine on a MOS chart and still lose the customer?

Book a demo and see how Cekura simulates those flows before release, scores voice clarity from the audio channel in production, and reruns your suite on every prompt and model change.

Frequently asked questions

What is a good Mean Opinion Score?

A MOS at or above 4.0 maps to satisfied users in the ITU-T G.107 planning table, and 3.6 sits where some users report dissatisfaction. Those bands assume 3.1 kHz handset telephony, so set your own thresholds per workflow and validate them against call outcomes.

Is MOS subjective or objective?

MOS is both subjective and objective, depending on which MOS you mean. ITU-T P.800.1 defines subjective scores from human panels, objective scores from perceptual models like POLQA, and estimated scores from planning models like the G.107 E-model, each with its own identifier.

What is the difference between MOS and POLQA?

The main difference between MOS and POLQA is that MOS is the score and POLQA is one way to produce it. POLQA, standardized as ITU-T P.863, compares degraded audio against a clean reference and outputs a predicted listening quality score written as MOS-LQO.

Can you measure MOS on a live production call?

Yes, you can measure MOS on a live production call using non-intrusive models such as DNSMOS or NISQA, which score the degraded audio on its own. Full-reference models like POLQA cannot run on live calls, because they need the clean original signal for comparison.

Is PESQ still a valid speech quality standard?

No, PESQ is no longer current. The ITU deleted the P.862 family from its catalog on January 5, 2024 and directs users to P.863, though many tools still ship PESQ implementations under the original name.

Does MOS work for testing AI voice agents?

Partly. MOS works for the audio layer of an AI voice agent and stops there. It cannot tell you whether the agent transcribed a callback number correctly, handled an interruption, called the right tool, or completed the caller's task, so pair it with conversation-level metrics.