A de-jitter buffer sized to absorb 50 ms of packet delay variation adds about 25 ms to every call on average, because ITU-T planning rules assume a buffer contributes half of its peak delay (ITU-T G.114). That 25 ms comes out of a budget that is already tight: conversational voice targets under 150 ms of one-way delay, with 400 ms as the limit (ITU-T G.1010). 3GPP now requires mobile telephony jitter buffers to keep jitter-induced losses below 1% on six standard network profiles (3GPP TS 26.114). Measurement has changed as well: ITU-T deleted the PESQ speech quality model on 5 January 2024 and points users to POLQA, whose fullband scale tops out at 4.8 rather than 4.5 (ITU-T P.862 page; ITU-T P.863.1). We aggregated data from ITU-T Recommendations G.114, G.107, G.107.1, G.1010, Y.1541, P.800, P.863 and P.863.1, IETF RFCs 3550, 3611 and 6716, 3GPP TS 26.114, Arcep’s 2025 mobile quality campaign, Microsoft’s Teams Call Quality Dashboard documentation, and peer-reviewed studies in Acta Acustica and the EURASIP Journal on Wireless Communications and Networking. For the market side of internet calling, see our VoIP and UCaaS statistics.
TL;DR
- One-way mouth-to-ear delay under 150 ms gives essentially transparent interactivity; 400 ms is the ceiling for general network planning (ITU-T G.114 2003).
- A de-jitter buffer adds half its peak delay on average: 25 ms for a buffer covering 50 ms of delay variation (ITU-T G.114 2003).
- Networks for interactive VoIP may vary packet delay by up to 50 ms, but the end-user target for conversational voice is under 1 ms (ITU-T Y.1541 2011; ITU-T G.1010 2001).
- 3GPP jitter buffers must keep jitter-induced loss below 1% on all six test profiles (3GPP TS 26.114, Release 19).
- Teams marks an audio stream Poor when average jitter exceeds 30 ms (Microsoft Learn, CQD stream classification 2026).
- Dropping from R 90 to R 70 cuts the share of users rating a call good or better from at least 97% to at least 73% (ITU-T G.107 2015).
- The wideband E-model runs from R 0 to 129, about 29% beyond the narrowband scale (ITU-T G.107.1 2019).
- ITU-T deleted PESQ (P.862) on 5 January 2024 (ITU-T P.862 page 2024).
- POLQA fullband: EVS at 48 kbit/s averages 4.70, G.711 only 3.6 (ITU-T P.863.1 2019).
- Playout shifts of 30 ms or less had a negligible effect on listeners, but PESQ misjudged them (Počta, Melvin and Hines, Acta Acustica 2015).
- Orange held 91% of calls at perfect quality nationally, Free Mobile 84% (Arcep 2025).
1. The Delay Budget: 150 ms for Natural Conversation, 400 ms as the Ceiling
Delay is a budget, not a single number. The codec, packetization, the network, the jitter buffer and the playback device all spend from the same pool, and the listener only feels the total. The ITU-T planning thresholds have not moved in two decades because they describe human conversation, not technology: below 150 ms one-way, most applications get essentially transparent interactivity; above 400 ms, ITU-T treats delay as unacceptable for general network planning.
What stands out is how little of that budget the network is supposed to use. The highest-interaction VoIP class in ITU-T Y.1541 allows a mean IP transfer delay of only 100 ms, which leaves 50 ms (150 ms minus 100 ms) for everything at the two ends of the call if the call is to stay transparent. A single G.723.1 codec consumes 97.5 ms in ITU-T’s wirebound planning table, against 0.375 ms for G.711, so codec choice alone can decide whether a call stays under 150 ms. Every real-time stage in the audio path, from noise suppression to a voice changer for Discord, draws on the same mouth-to-ear budget, which is why low processing latency matters as much as sound quality in live calls. Planners check these limits with the E-model, which by default only starts charging a delay penalty once absolute one-way delay passes 100 ms. The thresholds themselves are in ITU-T G.114 and ITU-T G.1010.
| Metric | Value | Source |
|---|---|---|
| One-way delay: limit for essentially transparent interactivity (mouth to ear) vs recommended ceiling for general network planning | Below 150 ms vs 400 ms | ITU-T G.114 (05/2003) |
| End-user target for conversational voice, one-way delay | Under 150 ms preferred, under 400 ms limit (assumes adequate echo control) | ITU-T G.1010 (11/2001) |
| Mean IP transfer delay bound: QoS Class 0 (high-interaction VoIP) vs Class 1 (interactive VoIP) | 100 ms vs 400 ms | ITU-T Y.1541 (12/2011) |
| One-way delay from codec processing in wirebound planning (3 frames plus look-ahead): G.711 / G.729 / G.723.1 | 0.375 ms / 35 ms / 97.5 ms | ITU-T G.114, Appendix I |
| Opus codec algorithmic delay range | 5 ms to 65.2 ms | IETF RFC 6716 (2012) |
| Opus frame sizes | 2.5, 5, 10, 20, 40 or 60 ms | IETF RFC 6716 (2012) |
| Default absolute one-way delay below which the E-model applies no delay impairment | 100 ms | ITU-T G.107 (06/2015) |
Context note: “look-ahead” is the extra slice of audio a codec must wait for before it can encode the current frame. G.114 (2003), G.1010 (2001) and Y.1541 (2011) are old but are the editions in force on the ITU-T site in October 2026. Most recent available data: ITU-T, 2001 to 2015.
2. Jitter: How Much Delay Variation Networks Are Allowed
Jitter is the variation in how long packets take to arrive. A call with a long but steady delay can still sound clean; a call whose packets arrive at uneven intervals needs a buffer, or it will break up. ITU-T allows interactive VoIP networks up to 50 ms of packet delay variation, yet the end-user target for conversational voice is under 1 ms of delay variation (ITU-T Y.1541; ITU-T G.1010). The gap between those two numbers is the job description of the jitter buffer: networks are expected to deliver jitter, and receivers are expected to hide it.
How jitter is reported matters as much as how much there is. RTP receivers compute a running estimate in which each new packet moves the figure by only one sixteenth of the difference, a gain the IETF chose because it “gives a good noise reduction ratio while maintaining a reasonable rate of convergence” (IETF RFC 3550). The same RFC recommends that control reports use 5% of session bandwidth, with a fixed minimum interval of 5 seconds. Because of that smoothing and that reporting pace, a single short delay spike barely moves the jitter a monitoring tool sees. Radio links add their own variation: in a 2021 VoLTE lab study, delay averaged 57.5 ms in a 20 to 150 ms range, while jitter clustered just above 31 ms (Lipovac et al., EURASIP JWCN 2021). The same study treats a 10% radio block error rate as the optimal operating point for LTE voice, falling to about 1% after one retransmission.
| Metric | Value | Source |
|---|---|---|
| Packet delay variation bound (99.9th percentile delay minus minimum delay), QoS Classes 0 and 1 | 50 ms | ITU-T Y.1541 (12/2011) |
| Packet loss ratio bound, QoS Classes 0 and 1 | 1 in 1,000 (1 x 10^-3) | ITU-T Y.1541 (12/2011) |
| Delay variation target at the end user, conversational voice | Under 1 ms | ITU-T G.1010 (11/2001) |
| Gain of the standard RTP interarrival jitter estimator | 1/16 per packet | IETF RFC 3550 (2003) |
| Average jitter above which Teams marks an audio stream as Poor | Above 30 ms | Microsoft Learn, CQD stream classification (2026), vendor threshold |
| Mean one-way delay in a VoLTE lab test (measured range 20 to 150 ms) | 57.5 ms | Lipovac et al., EURASIP JWCN 2021 |
| Value around which RTP jitter accumulated in the same VoLTE test | Slightly above 31 ms | Lipovac et al., EURASIP JWCN 2021 |
| Round-trip time of one LTE radio retransmission (HARQ) on FDD links | 8 ms | Lipovac et al., EURASIP JWCN 2021 |
Context note: HARQ (hybrid automatic repeat request) is the radio layer’s own fast retransmission. Each retry fixes a lost block but arrives later, which is one way a mobile link turns errors into jitter. The Teams 30 ms threshold is Microsoft’s own monitoring rule, not an industry standard; Microsoft notes that a Poor flag does not mean the user actually heard a problem.
3. Jitter Buffers: The Delay You Pay to Remove Jitter
A jitter buffer holds packets back so they can be played at an even pace. Every millisecond of holding is added delay; every packet that arrives after its playout time is lost, even though the network delivered it. Standards bodies have to pick a side of that trade-off, and 3GPP put its hard limit on loss: a compliant 3GPP jitter buffer must keep jitter-induced losses below 1% on every one of six standard delay and loss profiles, while keeping at least 90% of speech frames at or below a reference algorithm’s delay distribution shifted by 60 ms (3GPP TS 26.114). The loss limit is absolute, while the delay limit is only relative to a reference, which tells implementers that spending a little extra delay is acceptable and losing audio is not.
Adaptive buffers resize themselves during silences, and one controlled listening study found that small resizes went unnoticed. Playout shifts of 30 ms or less inserted into silence periods had a negligible effect on subjective listening scores, but the PESQ model reported a significant degradation that people did not hear (Počta, Melvin and Hines, Acta Acustica 2015). For live monitoring, RTCP XR reports let a receiver publish its nominal, maximum and absolute maximum buffer delay in milliseconds, plus whether the buffer is adaptive or fixed (IETF RFC 3611).
| Metric | Value | Source |
|---|---|---|
| Average delay a de-jitter buffer adds, for planning (worked example: buffer for a 50 ms delay variation range) | Half of its peak delay (25 ms in the example) | ITU-T G.114 (05/2003) |
| Jitter-induced loss limit for 3GPP multimedia telephony buffers | Below 1% on every test channel | 3GPP TS 26.114 (Release 19), clause 8.2.3.2.3 |
| Delay criterion: share of speech frames at or below the reference algorithm’s delay distribution shifted by 60 ms | At least 90% | 3GPP TS 26.114 (Release 19), clause 8.2.3.2.2 |
| Standard test profiles for jitter buffer compliance: number, length and built-in packet loss | 6 profiles of 7,500 IP packets each, with packet loss from 0% (profile 1) to 5.9% (profile 5) | 3GPP TS 26.114 (Release 19), Table 8.1 |
| Size of playout shifts in silence that had a negligible effect on listener scores | 30 ms or less | Počta, Melvin and Hines, Acta Acustica 2015 |
| Largest jitter buffer delay an RTCP XR report can express | 65,535 ms (16-bit field) | IETF RFC 3611 (2003) |
Context note: profile 5 carries 2 speech frames per packet (15,000 frames in 7,500 packets), while the other five carry 1 frame per packet. The 3GPP figures were read in the Release 19 text of TS 26.114; the 3GPP portal listed version 20.1.1 (Release 20) as the latest upload on 25 September 2026. Most recent available data on the listening test: Počta, Melvin and Hines, 2015.
4. Scoring Conversations: MOS Scales and the E-model R-Factor
Two scoring systems run side by side. The mean opinion score (MOS) comes from people: listeners rate samples on the five-point absolute category rating scale of ITU-T P.800, from 5 (Excellent) to 1 (Bad), after hearing short groups of sentences that each fit a 2 to 3 second slot. The E-model’s R-factor comes from a calculation: it starts from a basic signal-to-noise rating, subtracts impairments for noise, delay, echo, codec and loss, and adds an advantage factor for convenience. With every parameter at its default, R lands at 93.2 (ITU-T G.107).
The satisfaction table in Annex B is where R becomes business-relevant. Dropping from R 90 to R 70 cuts the share of users who rate a call good or better from at least 97% to at least 73%, and at R 50 the floor is only 27% good or better, with up to 38% rating it poor or worse. R 80 still keeps at least 89% at good or better and R 60 at least 50%, so perceived quality falls off faster than the linear R scale suggests. Wideband calls get a wider ruler: the wideband E-model runs from 0 to 129, about 29% beyond the narrowband scale (ITU-T G.107.1).
| Metric | Value | Source |
|---|---|---|
| Listening-quality scale used for subjective MOS tests | 5 = Excellent, 4 = Good, 3 = Fair, 2 = Poor, 1 = Bad | ITU-T P.800 (08/1996) |
| Narrowband R-factor with every E-model parameter at its default value | 93.2 | ITU-T G.107 (06/2015) |
| R 90 lower limit: estimated MOS / users rating good or better / users rating poor or worse | 4.34 / at least 97% / 0% (very satisfied) | ITU-T G.107 (06/2015), Annex B |
| R 70 lower limit: estimated MOS / good or better / poor or worse | 3.60 / at least 73% / at most 6% (some users dissatisfied) | ITU-T G.107 (06/2015), Annex B |
| R 50 lower limit: estimated MOS / good or better / poor or worse | 2.58 / at least 27% / at most 38% (nearly all users dissatisfied) | ITU-T G.107 (06/2015), Annex B |
| Maximum advantage factor A (quality users trade for convenience): wired / mobile in a building / mobile in a vehicle / multi-hop satellite | 0 / 5 / 10 / 20 | ITU-T G.107 (06/2015), Table 2 |
| R-factor range in the wideband E-model | 0 to 129 | ITU-T G.107.1 (06/2019) |
| R value RTCP XR reports label as toll quality vs value at or below which calls are regarded as unusable | 94 vs 50 | IETF RFC 3611 (2003) |
Context note: “toll quality” means the quality of a traditional landline call. RFC 3611 encodes MOS values as integers from 10 to 50 (MOS times 10), with 127 meaning not available. Its 94 toll-quality label and the 93.2 E-model default use different conventions and should not be read as a contradiction.
5. From PESQ to POLQA: How Objective Models Score Speech
Listening panels are slow and expensive, so networks are scored by objective models that compare a clean reference recording with the degraded one and predict the MOS a panel would give. ITU-T deleted PESQ (P.862) and its companion recommendations P.862.1 to P.862.3 on 5 January 2024, directing users to POLQA (P.863) (ITU-T P.862). As recently as its 2019 edition, the POLQA application guide still told engineers that P.863 “does not deprecate” P.862 and recommended running both in parallel, so any test plan written before 2024 deserves a second look.
The bigger trap is the scale. POLQA covers narrowband (300 to 3,400 Hz) through fullband (20 to 20,000 Hz) audio, and its third edition (2018) dropped the initial 14 kHz low-pass filter to handle fullband codecs such as Opus and EVS (ITU-T P.863). The narrowband mode maxes out at 4.5 while the fullband mode maxes out at 4.8, and the same G.711 call scores 4.4 in one and 3.6 in the other (ITU-T P.863.1). A “MOS of 3.6” means nothing until you know which mode produced it. For how operators publish these scores from real drive tests, see our cellular voice quality (MOS) statistics.
| Metric | Value | Source |
|---|---|---|
| PESQ status at ITU-T | P.862, P.862.1, P.862.2 and P.862.3 deleted on 5 January 2024 | ITU-T P.862 page (2024) |
| Raw PESQ score range | -0.5 to 4.5 | Lipovac et al., EURASIP JWCN 2021 |
| POLQA audio bandwidth coverage | Narrowband 300 to 3,400 Hz up to fullband 20 to 20,000 Hz (super-wideband 50 to 14,000 Hz simulated by band-limiting) | ITU-T P.863 (03/2018) |
| Maximum POLQA score: narrowband mode vs fullband mode | 4.5 vs 4.8 (transparent fullband references should score at least 4.75) | ITU-T P.863.1 (06/2019) |
| G.711 with IRS handset filter, average score: PESQ narrowband / POLQA narrowband / POLQA fullband | 4.4 / 4.4 / 3.6 | ITU-T P.863.1 (06/2019), Table 6 |
| AMR-NB at 12.2 kbit/s, average score: PESQ narrowband / POLQA narrowband / POLQA fullband | 4.0 / 4.2 / 3.2 | ITU-T P.863.1 (06/2019), Table 6 |
| Average POLQA fullband score: EVS 48 kbit/s fullband vs EVS 24.4 kbit/s super-wideband vs AMR-WB 23.85 kbit/s | 4.70 vs 4.65 vs 4.0 | ITU-T P.863.1 (06/2019), Table 6 |
| Recommended POLQA test sample | 3 to 6 s of active speech, at most 12 s in total; 8 sentence pairs from 2 male and 2 female talkers | ITU-T P.863.1 (06/2019) |
Context note: the Table 6 values are averages, and individual samples may differ. In the 2015 Acta Acustica study on jitter buffer time shifts (see section 3), POLQA performed significantly better than PESQ at predicting what listeners heard.
6. Field Measurement: What Regulators and Call Dashboards Record
Lab models describe what should happen; field campaigns show what does. France’s regulator Arcep counts a call as “perfect” only if it is held for two minutes with no audible disturbance, a stricter test than whether the call connects. Its 2025 campaign, built on more than 1 million measurements across all services between June and August 2025 (Arcep, mobile quality of service in mainland France 2025), shows only 76% to 83% of calls reaching perfect quality in rural areas, against 89% to 96% in dense areas.
Two details matter for VoIP specifically. Calls placed through messaging apps (over-the-top, or OTT, voice that runs as data) scored below the operators’ own voice service on the same networks. And on high-speed TGV trains, only 56% to 74% of calls were set up and held for two minutes. Enterprise platforms apply similar logic to every call: besides its 30 ms jitter rule, Teams flags an audio stream as Poor above 500 ms round trip or 10% packet loss, and only grades streams with more than 500 packets (Microsoft Learn, stream classification in CQD). For the network technology behind carrier voice, see our VoLTE and VoNR deployment statistics.
| Metric | Value | Source |
|---|---|---|
| Calls of perfect quality (held 2 minutes, no audible disturbance), mainland France: Orange / Bouygues Telecom / SFR / Free Mobile | 91% / 88% / 87% / 84% | Arcep 2025 campaign |
| Same indicator in dense areas: Orange / Bouygues Telecom / SFR / Free Mobile | 95% / 96% / 92% / 89% | Arcep 2025 campaign |
| Same indicator in rural areas: Orange / Bouygues Telecom / SFR / Free Mobile | 83% / 77% / 79% / 76% | Arcep 2025 campaign |
| Calls made through messaging apps (OTT voice) held 2 minutes with no audible disturbance | Orange 82%, Free Mobile 80%, SFR 77%, Bouygues Telecom 77% | Arcep 2025 campaign |
| Calls set up and held for 2 minutes on high-speed TGV trains | 56% (SFR) to 74% (Orange) | Arcep 2025 campaign |
| Round-trip time above which Teams marks an audio stream as Poor | Above 500 ms | Microsoft Learn, CQD stream classification (2026), vendor threshold |
| Packet loss rate above which Teams marks an audio stream as Poor | Above 10% (0.1) | Microsoft Learn, CQD stream classification (2026), vendor threshold |
Context note: the TGV figure measures calls set up and held for two minutes; Arcep’s release does not state that the “no audible disturbance” condition applies to the transport results, so it is not comparable with the perfect-quality rows. Arcep does not publish POLQA or MOS values in this release.
Summary: VoIP Jitter Buffer and Speech Quality by the Numbers
| Metric | Value | Source |
|---|---|---|
| One-way delay for essentially transparent interactivity vs planning ceiling | Below 150 ms vs 400 ms | ITU-T G.114 (2003) |
| Average delay added by a de-jitter buffer covering a 50 ms delay variation range | 25 ms (half its peak delay) | ITU-T G.114 (2003) |
| Mean IP transfer delay bound for high-interaction VoIP (QoS Class 0) | 100 ms | ITU-T Y.1541 (2011) |
| Delay variation target at the end user, conversational voice | Under 1 ms | ITU-T G.1010 (2001) |
| Opus algorithmic delay | 5 ms to 65.2 ms | IETF RFC 6716 (2012) |
| Jitter-induced loss limit for 3GPP jitter buffers | Below 1% on all 6 test profiles | 3GPP TS 26.114 (Release 19) |
| Share of speech frames that must stay within the reference delay plus 60 ms | At least 90% | 3GPP TS 26.114 (Release 19) |
| Average jitter that flags a Teams audio stream as Poor | Above 30 ms | Microsoft Learn (2026) |
| Mean one-way delay in a VoLTE lab test | 57.5 ms | Lipovac et al., EURASIP JWCN 2021 |
| Playout shifts in silence with negligible effect on listeners | 30 ms or less | Počta, Melvin and Hines, Acta Acustica 2015 |
| Users rating a call good or better at R 90 vs R 70 vs R 50 | At least 97% vs 73% vs 27% | ITU-T G.107 (2015) |
| Wideband E-model R range | 0 to 129 | ITU-T G.107.1 (2019) |
| R value labeled toll quality in RTCP XR | 94 | IETF RFC 3611 (2003) |
| PESQ deleted by ITU-T | 5 January 2024 | ITU-T P.862 page (2024) |
| Maximum POLQA score, narrowband vs fullband mode | 4.5 vs 4.8 | ITU-T P.863.1 (2019) |
| Packet delay variation bound for interactive VoIP networks (QoS Classes 0 and 1) | 50 ms | ITU-T Y.1541 (2011) |
| Perfect-quality calls nationally, best vs weakest operator in France | 91% (Orange) vs 84% (Free Mobile) | Arcep 2025 campaign |
| OTT app calls held 2 minutes without audible disturbance in France | 77% to 82% | Arcep 2025 campaign |
Methodology and Sources
Every figure above was read in the primary document during research for this article, either on the publisher’s page or in the full text of the recommendation, RFC or paper. Vendor guidance (Microsoft) is labeled as such and used only for operational thresholds.
- ITU-T: G.114, One-way transmission time (05/2003, with VoIP Appendix II amendment), G.107, The E-model (06/2015), G.107.1, Wideband E-model (06/2019), G.1010, End-user multimedia QoS categories (11/2001), Y.1541, Network performance objectives for IP-based services (12/2011)
- ITU-T speech quality assessment: P.800 (08/1996), P.862 (PESQ) status page, P.863 (POLQA) (03/2018), P.863.1 application guide (06/2019)
- IETF: RFC 3550, RTP (July 2003), RFC 3611, RTCP Extended Reports (November 2003), RFC 6716, Opus codec (September 2012)
- 3GPP: TS 26.114, IMS Multimedia Telephony, media handling and interaction (Release 19 text read; version 20.1.1 listed as latest)
- Arcep: Mobile quality of service in mainland France, 2025 campaign (press release, 20 November 2025)
- Microsoft: Stream classification in Call Quality Dashboard (Microsoft Learn, updated June 2026; vendor thresholds)
- Počta, Melvin and Hines: An Analysis of the Impact of Playout Delay Adjustments introduced by VoIP Jitter Buffers on Listening Speech Quality (Acta Acustica united with Acustica, 2015)
- Lipovac, Lipovac, Grbavac and Obradović: Practical cross-layer testing of HARQ-induced delay variation on IP/RTP QoS and VoLTE QoE (EURASIP Journal on Wireless Communications and Networking, 2021)
- Data watch: most of the delay and scoring thresholds come from standards that are old but still the in-force editions (G.114 2003, G.1010 2001, Y.1541 2011, P.800 1996, RFC 3550 and RFC 3611 2003); they are most recent available data, not stale measurements. Sources conflict on PESQ: ITU-T P.863.1 (2019) said POLQA does not deprecate PESQ and recommended running both, but ITU-T deleted P.862 on 5 January 2024. The 3GPP jitter buffer criteria were read in the Release 19 text through a specification mirror and should be rechecked against Release 20 (version 20.1.1, uploaded 25 September 2026). Arcep gives the 2025 measurement period as June to August in the body of its release and late May to August in the annex. The Lipovac et al. VoLTE study states a 1% significance level in one section and refers to 0.1 in another; only its descriptive delay and jitter values are used here. The E-model’s MOS cap for R above 100 is not quoted because the extracted text of G.107 was garbled at that point. TRAI drive-test reports (scanned PDFs) and Ofcom’s Mobile Matters pages (HTTP 403) could not be read and were excluded. Our research found few open peer-reviewed listening tests on jitter buffer behavior after 2016, and the web search budget for this session ran out early, so newer studies may exist and will be checked at the next refresh.
Last updated: October 24, 2026. We update this roundup quarterly, and the next refresh is expected when Arcep publishes the results of its 2026 mainland mobile quality-of-service campaign.