Choppy audio, robotic voices and people talking over each other all come down to three network numbers: delay, jitter and packet loss. VoIP call quality is usually summarised as a MOS score, but the score is only a model built from those three inputs. Measure them directly and you can fix the cause instead of guessing, whether the calls run through your own PBX or a cloud phone system.
Short answer: Aim for one-way delay under 150 ms, jitter under 30 ms and packet loss under 1%. Those keep the E-model R-factor above 80 and MOS around 4.0 or better with G.711, provided the phones use packet-loss concealment and the loss is random rather than bursty; without concealment, 1% loss alone takes G.711 down to about 3.8 (see the worked example). Measure the path with mtr for loss and latency, iperf3 -u with small packets for jitter, and the RTCP statistics on the PBX during real calls; fix problems with QoS on the upload side and by keeping phones off congested Wi-Fi. (Reviewed November 2026.)
Table of Contents
The three targets
| Metric | Good | Noticeable | Poor | What users hear |
|---|---|---|---|---|
| One-way delay | under 150 ms | 150 to 300 ms | over 300 ms | Talking over each other, awkward pauses |
| Jitter | under 30 ms | 30 to 50 ms | over 50 ms | Choppy or robotic audio |
| Packet loss | under 1% | 1 to 3% | over 3% | Missing syllables, dropouts |
The 150 ms delay figure comes from ITU-T G.114. Jitter buffers in phones hide some variation, but they do it by adding delay, so a jittery line also becomes a laggy one.
MOS and R-factor
MOS (Mean Opinion Score) rates quality from 1 (bad) to 5 (excellent). Monitoring tools estimate it with the ITU-T G.107 E-model, which starts from a perfect R-factor, subtracts penalties for delay, loss and the codec, and converts the result to MOS. As a rule of thumb, R 90 or above (MOS about 4.3) is what users call excellent, 80 to 90 (MOS 4.0 to 4.3) is good, 70 to 80 is where some users complain, and below 60 most people are unhappy. G.711 on a clean network tops out around MOS 4.4; compressed codecs such as G.729 start lower.
Worked example: from measurements to MOS
The full E-model has many inputs, but if you leave the rest at their G.107 defaults (which give R = 93.2) and assume echo is under control, three things decide the score: mouth-to-ear delay (Ta), the codec’s planning values from ITU-T G.113, and packet loss measured after the jitter buffer. Take a G.711 call with packet loss concealment (PLC), Ta of 150 ms (one-way network delay plus jitter buffer plus packetisation) and 1% random loss:
# 1. Delay impairment Idd (G.107 default class: mT = 100 ms, sT = 1)
# X = log2(Ta / 100) = log2(1.5) = 0.585
# Idd = 25 * ((1 + X^6)^(1/6) - 3 * (1 + (X/3)^6)^(1/6) + 2) = 0.16
#
# 2. Codec and loss impairment (G.113: G.711 with PLC, Ie = 0, Bpl = 25.1)
# Ie_eff = Ie + (95 - Ie) * Ppl / (Ppl / BurstR + Bpl)
# = 0 + 95 * 1 / (1 / 1 + 25.1) = 3.64
#
# 3. R = 93.2 - Idd - Ie_eff = 93.2 - 0.16 - 3.64 = 89.4
#
# 4. MOS = 1 + 0.035 * R + R * (R - 60) * (100 - R) * 0.000007
# = 1 + 3.129 + 0.195 = 4.32
Changing one input at a time shows what really moves the score (150 ms delay unless stated):
| Scenario | R | MOS |
|---|---|---|
| G.711 with PLC, 1% loss (above) | 89.4 | 4.32 |
| G.711 with PLC, 3% loss | 82.9 | 4.13 |
| G.711 with PLC, 1% loss, 300 ms delay | 74.8 | 3.81 |
| G.711 without PLC, 1% loss | 75.1 | 3.83 |
| G.711 without PLC, 2% random loss | 62.9 | 3.25 |
| G.711 without PLC, 2% bursty loss (BurstR 2) | 57.2 | 2.95 |
| G.729A with VAD, 1% loss | 77.8 | 3.94 |
Three lessons follow. Delay costs almost nothing until about 200 ms (Idd is about 3 at 200 ms and 15 at 300 ms). Loss hurts far more when the phone has no concealment. And bursty loss is worse than the same average spread evenly. G.113 lists these values for 10 ms G.711 packets and 20 ms G.729A packets, so treat the results as planning estimates, not measurements.
Measure the path
Run these from the office towards the PBX or provider during business hours, not at night when the line is idle:
# latency and loss per hop over 5 minutes
mtr -rwzbc 300 sip.provider.example
# jitter and loss with voice-sized UDP packets (needs an iperf3 server at the far end)
iperf3 -c 203.0.113.50 -u -b 100k -l 200 -t 60
# on Asterisk, live per-call RTCP statistics
asterisk -rx "pjsip show channelstats"
Loss that appears only on the first hop, and not at the destination, is usually the local router rate-limiting its replies to mtr, the same as on a middle hop. Blame the router or Wi-Fi only when the loss carries through to the final hop, or when the call’s RTP statistics or the router’s interface counters (drops and errors) show it too. Loss that starts mid-path and continues to the end points to the ISP. Loss on a single middle hop that does not carry through is just that router de-prioritising replies to mtr and can be ignored.
Reading Asterisk call statistics
pjsip show channelstats prints Receive and Transmit groups, each with Count, Lost, Pct and Jitter, followed by RTT. Three details catch people out:
- Units: Jitter and RTT are in seconds, so 0.012 means 12 ms. Pct is a whole-number loss percentage.
- Transmit loss comes from the far end: it is what the other side reports in its RTCP receiver reports. If the far end sends no RTCP, it stays at zero and tells you nothing.
- Raw RTCP jitter is in timestamp units: in a pcap or an RTCP report, RFC 3550 jitter is counted in RTP clock ticks. Divide by the clock rate: 8,000 for G.711 and G.729, and also for G.722 (RFC 3551); 48,000 for Opus (RFC 7587). A jitter of 240 on a G.711 call is 240 / 8000 = 0.03 s, or 30 ms.
For per-call logging, the dialplan function CHANNEL(rtcp,all) returns a semicolon-separated summary with lost packets, jitter, round-trip time and Asterisk’s Media Experience Score (MES), and CHANNEL(rtcp,all_mes) gives the minimum, maximum and average MES. MES is a 0 to 100 score Asterisk derives from G.107-style formulas; it is not a MOS. Phones that send RTCP XR (RFC 3611) can also report an R-factor and MOS directly; the MOS fields are sent as MOS × 10, so 41 means 4.1, and 127 means “not available”.
Fix the common causes
- Upload congestion: mark RTP with DSCP EF and SIP with CS3, give them strict priority on the edge router, and shape total upload to about 90% of the real line rate so queueing happens where QoS can act.
- Wi-Fi: put desk phones on wired Ethernet with PoE; on wireless use 5 GHz, WMM enabled and roaming tuned. Bluetooth headsets add their own delay.
- Too many hops: choose a provider or PBX location close to the users; each 1,000 km adds roughly 5 ms of one-way delay in fibre, more with routing detours.
- VPN overhead: tunnels add delay and fragmentation risk; check MTU and consider sending voice outside the tunnel over SRTP.
Common measurement mistakes
- Treating round-trip time as one-way delay: mtr and ping show round trips. Halve the figure only as a rough estimate on a symmetric path, then add packetisation and jitter-buffer delay to get mouth-to-ear delay.
- Ignoring late packets: G.113 notes that the loss that matters is measured at the output of the jitter buffer. Packets that arrive too late to play count as lost even though the network delivered them; RTCP XR reports them separately as the discard rate.
- Trusting averages: 1% loss over a call can still sound bad if it arrives in bursts, as the bursty row in the worked example shows.
- Testing one direction: iperf3 sends from client to server by default. Run it again with
-Rso the server sends, because upload and download problems are often different.
VoIP call quality at a glance


Official documentation: ITU-T G.107 E-model, ITU-T G.114 one-way delay, mtr manual, RFC 3550: RTP and RTCP.
Related guides: VoIP codec bandwidth · VoIP one-way audio fix · VPN MTU fragmentation.
Different symptom? See Common VoIP problems and how to fix them for the 11 problems we see most, with a free tool for each.
Frequently asked questions
What MOS score is acceptable for business calls?
A MOS of 4.0 or higher is good for business use and 3.6 to 4.0 is acceptable. Below about 3.5 users notice problems and complaints start.
Can a faster internet plan fix bad call quality?
Only if the line is actually saturated. Most quality problems come from jitter and queueing during uploads, which QoS and traffic shaping fix more reliably than extra bandwidth.
How often should call quality be monitored?
Continuously if possible. Most PBXs and cloud phone systems keep per-call quality statistics; review the worst calls weekly and correlate them with time of day and location.
What is the difference between MOS-LQ and MOS-CQ?
MOS-LQ (listening quality) leaves out delay; MOS-CQ (conversational quality) includes it and is usually derived from the E-model R-factor. A long-delay call can score well on MOS-LQ and lower on MOS-CQ, so check which one a tool reports.
Why does pjsip show channelstats show jitter as 0.004?
Asterisk prints jitter and round-trip time in seconds, so 0.004 is 4 ms, well inside the 30 ms target.
