The basics
Why measure speech quality?
Codecs, packet loss, jitter, transcoding between networks, echo, and level problems all degrade a call without showing up in classic network counters.
Speech quality measurement puts the listener's experience into a reproducible number that can be compared across networks and technologies and is suitable for SLAs.
From listening panels to algorithms
The reference is human judgement. In a listening test according to ITU-T P.800, a panel of listeners rates speech samples on a five-point scale, and the average is the Mean Opinion Score (MOS).
| MOS | Quality | Listening effort |
|---|---|---|
| 5 | Excellent | Complete relaxation possible |
| 4 | Good | Attention necessary, no appreciable effort |
| 3 | Fair | Moderate effort |
| 2 | Poor | Considerable effort |
| 1 | Bad | No meaning understood with any feasible effort |
Listening panels are slow and expensive. Objective models predict the MOS a panel would give, automatically and around the clock. The result is called MOS-LQO (Listening Quality Objective).
Full-reference measurement
The most accurate approach is intrusive (full-reference) testing: a known reference speech sample is sent through the connection, recorded at the far end, and the degraded recording is compared with the original using a perceptual model of human hearing and cognition. This is what PESQ® and POLQA® do.
-
Reference sample
A known speech sample is sent.
-
Through the connection
Across the path a customer's call takes.
-
Recorded at the far end
The degraded recording arrives.
-
Compared with the original
A perceptual model of hearing and cognition predicts the MOS-LQO.
Other approaches complement it:
- Parametric planning models
- Models such as the ITU-T G.107 E-model estimate quality (R-factor) from network parameters like delay, loss, and codec. Useful for planning and VoIP monitoring; they do not analyse the speech signal itself.
- Non-intrusive (no-reference) models
- These judge live traffic without a reference signal. They need no test calls, but are less precise.
PESQ®
- Standardised in 2001 as ITU-T P.862, for narrowband telephony (300 to 3400 Hz).
- P.862.1 maps the raw score to the MOS-LQO scale; P.862.2 extends PESQ® to wideband speech.
- For many years the industry standard for benchmarking fixed, mobile, and VoIP networks.
- Limits: designed before HD voice and modern codecs; less reliable with time-warping, packet-loss concealment, and wideband or super-wideband signals.
- Withdrawn by ITU-T in January 2024 in favour of P.863 (POLQA®). Because a large share of historical data is based on PESQ®, it remains in use for long-term comparisons.
POLQA®
- Standardised in 2011 as ITU-T P.863, the successor of PESQ®.
- Covers narrowband up to super-wideband / fullband speech, which makes it the reference for HD voice, VoLTE, VoWiFi, and modern VoIP codecs.
- Handles effects that confuse PESQ®: time stretching and compression, variable delay, and modern packet-loss concealment.
- Since the withdrawal of P.862, the ITU-T's full-reference model for listening quality, recommended for all new measurement campaigns.
AnchorMOS
x-fabric's speech quality score
AnchorMOS is x-fabric's own method for scoring the speech quality of HD voice calls. Like PESQ® and POLQA®, it is a full-reference measurement in which a known speech sample is sent through the connection and the recording from the far end is compared with the original. The result is a MOS on the familiar five-point scale.
- Developed in-house, without depending on a patented scoring algorithm.
- Tuned to real network conditions rather than laboratory conditions.
- On 997 of 1,000 test calls, AnchorMOS lies less than a tenth of a point from the reference method.
- Can run inside the mosprobe analysis chain alongside PESQ® and POLQA®, allowing direct comparison of the results.
What a good measurement needs
Questions and answers
Reading the scores
What is a good MOS?
It depends on the model and its mode, so a score is meaningful only when both are known. As a guide for narrowband telephony: with POLQA® in narrowband mode, an undisturbed G.711 call scores about 4.3 and a mobile call with AMR at 12.2 kbit/s about 4.2. The ITU-T quality classes for network planning (G.109) count users as satisfied from about 4.0; between 3.6 and 4.0 some are dissatisfied, below 3.6 many.
In super-wideband or fullband mode the same calls score lower, because the missing bandwidth counts as a loss. There, transparent transmission reaches 4.75.
Why does no measurement reach a MOS of 5?
Because listeners rarely award a 5 either. Even an undistorted recording averages around 4.5 in a listening test, since not every listener gives it the top mark. Objective models are trained to predict these panel scores, so they share the ceiling: PESQ® and POLQA® in narrowband mode end at about 4.5, POLQA® in super-wideband mode at 4.75. A narrowband score of 4.4 is therefore close to the best a call can reach.
Can PESQ® and POLQA® scores be compared?
Not directly. They are different models, validated on different data, and they react differently to modern codecs, time warping and packet-loss concealment. Even POLQA® itself scores on two scales: the same G.711 call reaches about 4.3 in narrowband mode and about 3.5 in super-wideband mode.
Every result should therefore name the model and the mode. When switching from PESQ® to POLQA®, measure both for a while to connect the old series to the new one.
What do MOS-LQS, MOS-LQO and MOS-CQE mean?
ITU-T P.800.1 labels every MOS by what was rated and where the number comes from. LQ stands for listening quality, CQ for conversational quality, which also takes delay and echo into account. The last letter names the source: S for a subjective test with listeners, O for an objective model that analyses the signal, such as PESQ® or POLQA®, and E for an estimate from network parameters, such as the E-model. A POLQA® result is a MOS-LQO, an E-model result a MOS-CQE.
Choosing the method
PESQ® or POLQA®: which method for what?
For new measurements, POLQA® (ITU-T P.863): ITU-T withdrew PESQ® (P.862) in January 2024 and refers to P.863 instead. POLQA® covers narrowband to fullband speech, handles modern codecs such as AMR-WB, EVS and Opus, and copes with time warping and variable delay; its current edition from 2018 analyses the full audio band up to 20 kHz. PESQ® remains useful wherever long PESQ®-based series must stay comparable, mainly in narrowband networks.
What do narrowband, wideband, super-wideband and fullband mean?
They describe how much of the audio spectrum a call carries. The wider the band, the more natural and intelligible the voice.
| Band | Frequencies | Typical codecs |
|---|---|---|
| Narrowbandclassic telephony | 300–3400 Hz | G.711, AMR-NB |
| WidebandHD voice | 50–7000 Hz | G.722, AMR-WB |
| Super-wideband | 50–14000 Hz | EVS, Opus |
| Fullband | 20–20000 Hz | EVS, Opus |
Narrowband is what analogue lines, ISDN and older mobile networks carry. HD voice usually means wideband; EVS in VoLTE goes further.
How do you measure HD voice, VoLTE and VoWiFi?
With POLQA® in fullband or super-wideband mode, not in narrowband mode, whose scale does not reward the wider band, and with reference speech recorded in full bandwidth and sampled at 48 kHz. Measure end to end, including the interconnects: if a call is transcoded to narrowband anywhere on the way, the HD advantage is gone, and the score shows it. Typical published values in super-wideband mode are about 3.8 for an AMR-WB call and about 3.5 for G.711.
VoLTE and VoWiFi use the same IMS voice codecs, AMR-WB and EVS, so both are measured the same way, in both directions.
When is a measurement without a reference signal enough?
When the aim is to watch large volumes of real calls rather than to benchmark. No-reference models such as ITU-T P.563, made for narrowband, judge a received signal on its own; parametric models estimate quality from packet data such as loss, jitter and codec, without listening at all. Neither needs test calls, but both are less precise than a full-reference measurement and cannot tell whether what arrived is what was said.
How do the R-factor and the MOS relate?
The R-factor is the result of the ITU-T G.107 E-model, a planning tool that subtracts impairments from delay, echo, packet loss and the codec from a maximum of 100 for narrowband. G.107 converts R into an estimated MOS (MOS-CQE): R 80 corresponds to about 4.0, R 70 to 3.6 and R 50 to 2.6. An unimpaired narrowband connection reaches R 93.2, about MOS 4.4; the wideband E-model (G.107.1) extends the scale to 129.
In practice
Which speech samples should be used?
Use natural recorded speech rather than tones or synthetic signals: short sentences from female and male talkers, with natural pauses, at a defined level. ITU-T P.501 provides such speech files in several languages, free to download, and the application guide ITU-T P.863.1 explains how much material to use and which levels to set.
Changing the samples changes the scores, so a campaign keeps the same set throughout.
Why are good network values such as packet loss and jitter not enough?
Because they describe packet behaviour rather than how a call sounds. A jitter buffer absorbs jitter at the cost of delay, packet-loss concealment hides losses, and the same loss rate sounds different depending on the codec and on whether packets go missing singly or in bursts. Transcoding, echo, wrong levels, clipping, noise suppression and one-way audio do not show in packet statistics at all.
What about delay and echo?
Listening-quality models such as PESQ® and POLQA® align the recording with the reference in time and rate only what is heard. They do not penalise delay, and the talker's own echo never reaches the recording at the far end, so both have to be measured separately.
ITU-T G.114 states that most applications experience essentially transparent interactivity below 150 ms of one-way, mouth-to-ear delay, and that 400 ms should not be exceeded in network planning. Echo becomes more disturbing the longer the delay.