MOS: definition and a practical example
Short answer: MOS (Mean Opinion Score) is the 1-to-5 score, defined by ITU-T P.800, that expresses the voice quality perceived by someone listening to a call.
In Portuguese: MOS
What is MOS?
MOS (Mean Opinion Score) is a 1-to-5 score that sums up how people perceive the quality of a voice call. The scale comes from ITU-T P.800: 5 is excellent, 4 good, 3 fair, 2 poor and 1 bad. Originally, the score came from tests with real listeners who rated audio samples; the average of those ratings is the MOS.
| Score | Quality (ITU-T P.800) |
|---|---|
| 5 | Excellent |
| 4 | Good |
| 3 | Fair |
| 2 | Poor |
| 1 | Bad |
Nobody convenes a listening panel for every call. That is why ITU-T P.800.1 separates the types of MOS by where the score comes from: MOS-LQS is the subjective listening score, given by people; MOS-LQO is the objective score, calculated by an algorithm that compares the received audio with the original; MOS-LQE is a listening estimate; and MOS-CQE is the conversational estimate that comes out of a network model. When a report shows “MOS 4.1”, the first question is: which MOS? A score estimated from network numbers and a score measured on the audio can diverge sharply on the same call.
R-factor and the E-model
The E-model, from ITU-T G.107, is a network planning model. It starts with an R-factor and subtracts penalties for delay, packet loss, codec and echo. With the default parameters, R comes out around 93.2. The standard provides the formula that converts R to MOS:
| R-factor | MOS-CQE |
|---|---|
| 90 | 4.34 |
| 80 | 4.03 |
| 70 | 3.60 |
| 60 | 3.10 |
| 50 | 2.58 |
MOS-CQE
MOS-CQE is the estimated conversational score, the E-model’s output expressed on the MOS scale. The maximum is 4.5. It never listens to the audio: it calculates the score from parameters such as delay, loss and codec. It works well for sizing a network, but it cannot notice a silent call, one side with no audio or a recorded carrier message in place of the person being called.
In practice
Suppose a network device reports R = 80 for a call. By the G.107 table, that equals a MOS-CQE of 4.03, a “good” call. Now suppose that call landed on a carrier recording saying the number is not available. The packets arrived in order, with no loss and little jitter, so the E-model keeps showing 4.03. For the caller, the call failed.
The reverse happens too. A network with 0% loss can deliver audio chopped up by a misconfigured transcoder. The E-model does not see the problem, because it lives in the content of the audio, not in the packets. Only a score measured on the received sound, compared with the expected sound, reveals both cases.
How Monitor uses this
Monitor never estimates MOS from network numbers. It measures MOS-LQO on the audio actually received on test calls. The details of how the score is calculated are in voice quality with MOS-LQO and in the ViSQOL entry.
Frequently asked questions
What is a good MOS for VoIP?
On the ITU-T P.800 scale, 4 means good and 3 means fair. The E-model's MOS-CQE tops out at 4.5, so a score above 4 is already close to the best the model allows. Scores measured on the audio, such as MOS-LQO, depend on the algorithm and the codec's bandwidth, so it only makes sense to compare scores of the same type.
Are MOS and R-factor the same thing?
No. The R-factor is the output of the E-model (ITU-T G.107), on a scale that reaches about 93.2 with the default parameters. It is converted to MOS-CQE with a formula from the standard itself: R 80 equals 4.03 and R 70 equals 3.60. The R-factor is a network planning estimate, not a score taken from the audio.