What is a good MOS score for VoIP calls?

4.0 and above, good

Callers are satisfied and few would remark on the quality. For G.711 the realistic best is ~4.3 to 4.4. Wideband codecs (G.722, Opus) can score higher because they carry more of the voice.

3.6 to 4.0, fair

Usable, but some callers notice, occasional clipping or a slightly synthetic edge. Worth watching, especially if a route trends downward.

Below 3.6, poor

Audible problems, gaps, robotic audio, talking over delay. Complaints and repeat calls follow, and on a wholesale route this is where minutes get disputed.

MOS is a summary of network conditions plus the codec. Three transport factors move it:

• Packet loss, audio that never arrived. Keep it under about 1 percent. Past a few percent, speech breaks up.
• Jitter, packets arriving unevenly. The jitter buffer absorbs some, then starts discarding. Keep jitter under about 30 milliseconds.
• Latency and one-way delay, which makes conversation feel like a walkie-talkie once round-trip passes roughly 300 milliseconds.

The codec sets the ceiling. G.711 is clear but narrowband, so its ceiling is about 4.4 on the scale. A wideband codec can sound better but costs more bandwidth. So a "low" MOS is only meaningful against the codec's own ceiling, not against a flat 5.0.

A measured MOS (ITU-T P.800) comes from people rating audio, or from a perceptual algorithm that compares the original and received audio sample by sample. It is the gold standard, but it needs reference audio and does not run on live traffic.

An estimated MOS (the ITU-T G.107 E-model) is computed from the network conditions and the codec: take the packet loss and jitter, derive a transmission-quality rating, map it to a 1 to 5 score. It needs no reference audio, so it is what almost every live VoIP monitor reports. It is an estimate of transport quality, not a perceptual measurement, and tooling should label it that way.

Most quality monitoring relies on synthetic test calls or dedicated probes: you learn about a bad route from a robot that dialled it at 3 a.m., not from the real calls your customers made. The media stream itself already carries everything MOS needs. Any device on the media path can read the RTP, recover packet loss from the sequence numbers, measure jitter from arrival timing, and run the E-model, per leg, on real calls.

That is the approach Open Voice Shield takes. It sits as one extra hop on your SIP path and already relays and records the media to analyze each call for fraud. The same capture yields an estimated MOS, jitter and packet loss for every analyzed call, split into the caller and callee legs, with no probes, synthetic calls or installed software. Because the caller and callee legs are scored separately, a route where the far end degrades shows up as a low callee-leg MOS while your own side stays clean, which points straight at the vendor. The metrics ride the same record as the fraud verdict and are quoted in the alert email, and the field reference is in the documentation.

Above 4.0 is good, 3.6 to 4.0 is fair, below 3.6 is poor, and remember the G.711 ceiling of about 4.4. Keep packet loss under 1 percent and jitter under 30 milliseconds and the score follows. If something already carries your media, it can report the estimated MOS of every real call you make, with no lab needed.