मुख्य सामग्री पर जाएँ
तकनीकी 7 नवंबर 2026 11 मिनट का पठन

How to Measure VoIP Call Quality (MOS, Jitter)

Measure VoIP call quality the way the standards do: one-way delay against ITU-T G.114, jitter and loss from RTP, MOS from the E-model, and where to measure.

D
7 नवंबर 2026

त्वरित उत्तर

Measure four things on every call: one-way delay, jitter, packet loss and an estimated MOS. ITU-T G.114 sets 150 ms mouth-to-ear as the level below which most uses feel transparent, and 400 ms as the upper planning limit. RTP and RTCP define jitter and loss. The E-model (ITU-T G.107) turns the measurements into a score. Measure at the agent, the media server and the carrier, and read distributions, not averages.

"The calls sound bad" is the hardest complaint in a contact centre, because it is true for some calls, some agents and some hours, and false for the rest. The only way to deal with it is to measure each call, in the same way, at known places, and compare the results with published targets rather than rules of thumb.

This guide uses the standards themselves: ITU-T G.114 for delay, RFC 3550 for jitter and loss, and ITU-T G.107 for the model behind estimated MOS. It explains where to measure on a typical contact-centre call, why averages mislead, and how to test a new carrier route before you trust it with a campaign.

What to measure

Measureयह आपको क्या बताता हैWhere it comes from
One-way delayHow long speech takes from mouth to ear; drives talk-over and awkward pausesNetwork path plus codec, packetisation and jitter buffer
JitterHow uneven packet arrival is; drives choppy audioRTP and RTCP (RFC 3550)
Packet lossHow many audio packets never arrived; drives gaps and robotic soundRTCP reception reports
MOS (estimated)One number summarising the above for a quick comparisonA model such as the E-model, or listening tests

Keep the terms apart. Jitter और राज्यमंत्री have their own glossary entries; post-dial delay is a different delay entirely: it is the wait before ringing starts, and is not part of mouth-to-ear delay once the call is answered.

Delay: what G.114 actually says

ITU-T Recommendation G.114 covers one-way transmission time. Two statements from the 05/2003 text matter for a contact centre, and they are often paraphrased loosely:

  • "Although a few applications may be slightly affected by end-to-end (i.e., "mouth-to-ear" in the case of speech) delays of less than 150 ms, if delays can be kept below this figure, most applications, both speech and non-speech, will experience essentially transparent interactivity." (ITU-T G.114, section 4)
  • It is recommended not to exceed a one-way delay of 400 ms for general network planning. Delays above 400 ms are described as unacceptable for general planning purposes, although the text recognises exceptional cases such as an unavoidable double satellite hop for a hard-to-reach location. (ITU-T G.114, section 4)

Note what these are not. 150 ms is not a pass mark that guarantees good calls, and 400 ms is not a fail line for individual calls; it is a planning limit. G.114 also says the E-model should be used to estimate the effect of one-way delay, including all delay sources, on conversational speech quality. Delay is mouth-to-ear, so it includes more than the network:

  1. Capture and codec delay at the sender.
  2. Packetisation: with the default 20 ms packet interval for G.711 (RFC 3551), at least one packet interval is spent filling a packet.
  3. Network transit.
  4. The receiver's jitter buffer, which holds packets back to smooth arrival.
  5. Playout and any transcoding or media-server hop.

An illustrative delay budget

The figures below are invented for the arithmetic and not measured on any network. A call between an agent and a customer on another continent:

Component (illustrative)Delay
Packetisation (20 ms frames)20 ms
Agent to media server, network30 ms
Media server to carrier, network60 एमएस
Carrier to destination handset40 ms
Jitter buffers on both ends60 एमएस
Total210 ms

This route is above 150 ms and well below 400 ms. It will not be broken, but it will not be "essentially transparent" either, and the sheet shows where to look first: the long carrier leg and the jitter buffers. The point of the exercise is the shape of the sum, not the numbers.

Jitter and loss as RTP and RTCP define them

RFC 3550 defines both measurements precisely, which is why different tools can agree on them.

Interarrival jitter

For two packets, RFC 3550 defines D as the difference in their relative transit times: the gap between arrival times minus the gap between the sender's timestamps. Jitter is then smoothed continuously with a gain of 1/16, according to J = J + (|D| - J) / 16 (RFC 3550, section 6.4.1). Two practical consequences:

  • It is an estimate that moves slowly. One bad packet does not spike it, and a sudden burst of delay shows up in the number a little after it happened.
  • It is measured at the receiver in RTP timestamp units. Tools convert to milliseconds using the codec clock rate, so compare values from tools that make the same conversion.

Packet loss

RFC 3550 defines the cumulative number of packets lost as the packets expected minus the packets received, where "received" includes late or duplicate packets. As a result, late packets are not counted as lost, and the number can even go negative if there are duplicates (RFC 3550, section 6.4.1). This matters for audio: a packet that arrives too late for the jitter buffer is not "lost" by this count, but it is still missing from what the listener hears. The practical advice is to measure loss and also the discard rate of the jitter buffer, where your tool exposes it. Browsers do: the W3C WebRTC statistics include packetsDiscarded, the cumulative number of RTP packets the jitter buffer discarded because they arrived late or early and were not played out (W3C, WebRTC Statistics).

MOS: listening tests versus estimates

MOS stands for mean opinion score. A real MOS is the average of ratings given by people who listened to calls in a controlled test. That is the gold standard and also impractical to run on every call. Monitoring tools therefore compute an estimated MOS from network measurements using a model. The model the ITU-T recommends for this is the E-model in ITU-T G.107, which the Recommendation calls the common ITU-T transmission rating model; its in-force version is 06/15 and whose title is "The E-model: a computational model for use in transmission planning" (ITU-T G.107). The model's primary output is a transmission rating factor, R, from 0 to 100, and Annex B of G.107 turns R into an estimated mean opinion score for conversation on a 1 to 5 scale. G.107 itself stresses that such estimates are made only for transmission planning purposes and not for actual customer opinion prediction, and its scope is 3.1 kHz handset telephony, that is, narrowband calls (ITU-T G.107, section 1 and Annex B, text of 06/15).

Three cautions:

  • It is a planning model. It is good for comparing routes and spotting trends; it does not hear the call. Background noise at the agent's desk, a bad headset or a clipped microphone may not show up.
  • Different tools make different assumptions. Two tools can report different MOS for the same call. Compare like with like: one tool, same settings, over time.
  • Back it with ears. Pick a handful of calls from each score band and listen. If the model says 4.0 and the call sounds poor, trust the call and investigate what the model cannot see.

Where to measure

A contact-centre call has three legs, and each needs its own view. A single end-to-end number tells you something is wrong but not where.

1. The agent leg (browser or softphone)

For browser-based agents the W3C WebRTC statistics specification gives the measurements. The inbound audio statistics include jitter and packets lost, and also jitter buffer delay (W3C, WebRTC Statistics). The remote inbound statistics, which reflect what the far end reports about your outgoing audio, include jitter, packets lost and round-trip time. Round-trip time is the one to use for a first estimate of network delay on this leg. Half of it is a rough one-way figure, and only if the path is symmetrical.

Sample these about once a second during a call and keep the series. Poor audio on this leg usually points to the agent's Wi-Fi, headset or office network rather than the carrier.

2. The media server

The media server sees both sides. It can report loss and jitter on each leg separately, which is what lets you tell whether the problem was between the agent and the platform, or between the platform and the carrier. If your platform records its own per-call quality score, this is where it comes from.

3. The carrier leg

Measure what the carrier sends you: jitter and loss on the incoming media, plus the answer rate and call duration trends for the route. Tools on your side can only measure to your edge; for what happens beyond it, ask the carrier for their own statistics, and compare them with yours.

When all three legs look fine and the call is still poor, suspect the far end: a mobile handset in a weak signal area is outside everyone's control.

Why averages hide the calls that went bad

Quality is a tail problem. Suppose 97 calls in 100 score well and 3 are terrible. The average barely moves; the three customers and their agents remember. Better ways to read the data:

  • Percentiles. Look at the 5th percentile of MOS, or the 95th percentile of jitter.
  • A bad-call count. Calls below a threshold you choose, as a percentage of total calls, per day.
  • Splits. By carrier route, destination country, agent site, time of day and codec. A problem concentrated in one split has a cause you can find.
  • Correlation with outcomes. Do short calls and early hang-ups cluster in the bad-quality group?

Running a quality test on a new carrier route

  1. Define the target before you test. For example, an estimated MOS floor, a jitter ceiling and a loss ceiling that you accept for a campaign, written down in advance.
  2. Pick representative destinations. Test the mobile and fixed number ranges you will actually dial, not only one friendly number.
  3. Test at different hours. Peak-hour congestion does not show up in a morning test.
  4. Use enough calls. A handful of calls tells you little about the tail. Run a pilot of real calls at low volume, and watch the percentiles.
  5. Check each leg. Compare agent, media and carrier numbers so the blame lands in the right place.
  6. Listen. Review calls from the worst band and the median band.
  7. Compare routes. Send the same share of traffic through two routes and compare. The guide to choosing a SIP trunk provider for a dialer covers what else to compare.

DialerBee कैसे काम करता है

DialerBee captures the mean opinion score per call in its operations center, and low-MOS alerts surface on the wallboard while the shift is running. The पर्यवेक्षक उपकरण wallboard includes a MOS card for the audio quality of calls in progress, and the संचालन केंद्र is where the alerts and history sit. That gives a supervisor the per-call number and the live view; the work described above, setting targets, reading tails and splitting by leg and route, is what turns it into decisions.

Measurement checklist

  • Targets written down: delay (150 ms goal, 400 ms planning ceiling), jitter, loss, MOS floor.
  • Delay budget calculated for each major route, including packetisation and jitter buffers.
  • Jitter and loss captured on the agent, media and carrier legs.
  • Estimated MOS tracked per call with one consistent tool.
  • Percentiles and a bad-call count reported, not only averages.
  • Splits by route, destination, site and hour.
  • A listening review each week of calls from the worst band.
  • New routes piloted on real calls before full volume.

अक्सर पूछे जाने वाले प्रश्न

What is acceptable latency for VoIP?

ITU-T G.114 says that if one-way, mouth-to-ear delay can be kept below 150 ms, most applications, speech included, will experience essentially transparent interactivity. It also says delays above 400 ms are unacceptable for general network planning, though it accepts exceptions such as an unavoidable double satellite hop. Treat 150 ms as the target and 400 ms as the planning ceiling.

How do I measure jitter?

Use the interarrival jitter that RTP defines. For each packet, compare the gap between arrival times with the gap between sending timestamps, and smooth the absolute difference with a gain of 1/16. RTP receivers report it in RTCP, and browsers expose a jitter value in their WebRTC statistics. Read it per call and as a distribution, not only as an average.

How is packet loss measured?

RTCP reception reports carry a cumulative count of packets lost since the start of reception, defined as packets expected minus packets received. Late packets are not counted as lost, and duplicates can make the figure negative. Note that a packet which arrives too late for the jitter buffer is still audible as a gap even though RTP does not count it as lost.

What is MOS and who measures it?

MOS is the mean opinion score, a 1 to 5 rating of call quality. A true MOS comes from people scoring calls in a listening test. Monitoring tools usually report an estimated MOS calculated from network measurements using a model such as the ITU-T E-model (G.107). Treat the estimate as a consistent comparison tool, not a replacement for listening to calls.

Can I get one-way delay from a round-trip time?

Only approximately. Half the round-trip time is a common estimate and assumes the path is symmetrical, which is not always true. It also leaves out the codec, packetisation and jitter buffer delay that make up the rest of mouth-to-ear delay. Add those before you compare the result with the 150 ms target.

Why do average scores look fine when agents complain?

An average mixes thousands of good calls with the few bad ones. If three per cent of calls are bad, the average barely moves, but agents and customers remember the bad ones. Look at percentiles, a count of calls below a threshold, and the breakdown by carrier route, agent location and time of day.

सूत्रों का कहना है

  1. ITU-T G.114, One-way transmission time (05/2003 text, section 4: below 150 ms gives essentially transparent interactivity; delays above 400 ms are unacceptable for general network planning). itu.int/rec/T-REC-G.114 (7 नवंबर 2026 को जाँचा गया)
  2. ITU-T G.107, The E-model: a computational model for use in transmission planning (06/15; section 1 scope, Annex B quality measures derived from R). itu.int/rec/T-REC-G.107 (7 नवंबर 2026 को जाँचा गया)
  3. RFC 3550, RTP: A Transport Protocol for Real-Time Applications (section 6.4.1: interarrival jitter and cumulative packets lost). rfc-editor.org/rfc/rfc3550 (7 नवंबर 2026 को जाँचा गया)
  4. RFC 3551, RTP Profile for Audio and Video Conferences with Minimal Control (Table 1: 20 ms default packet interval for PCMU and PCMA). rfc-editor.org/rfc/rfc3551 (7 नवंबर 2026 को जाँचा गया)
  5. W3C, Identifiers for WebRTC's Statistics API, Candidate Recommendation Draft of 25 September 2025 (jitter, packets lost, packets discarded, round-trip time, jitter buffer delay). w3.org/TR/webrtc-stats (7 नवंबर 2026 को जाँचा गया)

संबंधित पठन

संबंधित लेख: How Much Bandwidth Does a Call Center Need? · What Is VoIP? Voice over IP Explained · What Is VoIP QoS? Quality of Service

DialerBee को काम करते हुए देखने के लिए तैयार हैं?

15 मिनट के लाइव डेमो का अनुरोध करें, या 14 दिन का फ़्री ट्रायल शुरू करें — न स्लाइड, न कोई प्रतिबद्धता। कॉलिंग तब शुरू होती है जब आपका अपना SIP ट्रंक या PBX जुड़ जाए और मंज़ूर हो जाए।

14 दिन का फ़्री ट्रायल · मुफ़्त AI क्रेडिट · कोई कार्ड नहीं · 11 भाषाएँ · BYOC · अनुपालन में सहायक नियंत्रण

पूरी साइट अंग्रेज़ी में देखें →