Acceptable jitter for VoIP is under 30 ms: Cisco's Enterprise QoS design guide (version 3.3, 2005) targets average one-way jitter under 30 ms, and Microsoft asks for under 30 ms from a Teams client to its network edge in any 15-second interval. The ITU-T recommendations we read set no limit on RTP interarrival jitter: Y.1541 bounds network delay variation at 50 ms (99.9th percentile), and the under-1-ms target in G.1010 describes delay variation as the user perceives it after buffering. The 30 ms figure travels with 150 ms one-way delay (ITU-T G.114) and 1% loss (Cisco) and only makes sense next to them: the ITU-T G.107 E-model has no jitter input, so jitter reaches the listener only through the buffer, as added delay or as late packets thrown away.
What is jitter in VoIP, and how does RFC 3550 calculate it?
Jitter in VoIP is the variation in transit time between consecutive packets of one RTP stream. The sender spaces packets evenly (20 ms is the RFC 3551 default for PCMU), queues delay each packet differently, and they arrive unevenly.
RFC 3550 §6.4.1 defines the value every RTP receiver reports. With D(i,j) = (Rj − Ri) − (Sj − Si) for RTP timestamps S and arrival times R, the receiver updates J(i) = J(i−1) + (|D(i−1,i)| − J(i−1))/16 on every packet: "the mean deviation (smoothed absolute value)" of D.
- Units. RTCP carries J in timestamp units; divide by the clock rate, 8,000 Hz for PCMU (RFC 3551) or 48,000 Hz for Opus (RFC 7587). A reported 240 is 30 ms for PCMU and 5 ms for Opus.
- Smoothing. One late packet moves J by a sixteenth of its deviation: J tracks sustained variation, not the worst packet.
- Intended use. RFC 3550 §6.4.4 calls the field "only a snapshot of the jitter at the time of a report" that "is not intended to be taken quantitatively".
Jitter vs latency vs packet loss: what each one measures
Latency is how long a packet takes, jitter is how much that time varies from packet to packet, and packet loss is the share of packets that never play out.
| Metric | What it is | Not the same as |
|---|---|---|
| One-way delay | Mouth to ear, "including all delay sources" (G.114); the E-model's Ta | Network RTT: two directions, no codec or buffer |
| Round-trip time | A − LSR − DLSR from RTCP (RFC 3550); roundTripTime in getStats | Twice the one-way delay: "some links have very asymmetric delays" (RFC 3550) |
| Interarrival jitter | Smoothed variation of transit time, per direction (RFC 3550) | Y.1541 IPDV, a 99.9th-percentile spread; packet reorder, a separate Teams target |
| Packet loss | Packets that never arrived (RTCP fraction lost) | Jitter-buffer discards, counted separately in RFC 3611; both "have equal effect on call quality" |
Who publishes the 30 ms jitter, 150 ms delay and 1% loss limits?
Four vendor documents set jitter at 30 ms. ITU-T bounds network delay variation and post-buffer delay variation, not RTP jitter, and the Zoom Phone dashboard article gives no per-metric limit. The numbers cover different legs and intervals.
| Source | Jitter | Delay | Loss | Measured over |
|---|---|---|---|---|
| ITU-T G.114 (05/2003) | None | Below 150 ms, "essentially transparent interactivity" for most applications; 400 ms planning limit | Not covered | One way; 150 ms mouth-to-ear, 400 ms UNI to UNI |
| ITU-T G.1010 (11/2001), conversational voice | Delay variation < 1 ms, as perceived after buffering | < 150 ms preferred, < 400 ms limit, assuming adequate echo control | < 3% packet loss ratio, assuming concealment | End user |
| ITU-T Y.1541 (12/2011), Classes 0 and 1 (VoIP) | IPDV ≤ 50 ms: 99.9th percentile of delay minus the minimum | Mean ≤ 100 ms (Class 0), ≤ 400 ms (Class 1) | ≤ 1 × 10−3 (0.1%) | Network only, UNI to UNI; every minute |
| Cisco Enterprise QoS SRND 3.3 (2005) | Average one-way under 30 ms | ≤ 150 ms mouth-to-ear, extendable to 200 ms | ≤ 1% | End-to-end design target |
| Microsoft Teams, client to Microsoft edge | < 30 ms in any 15 s | < 50 ms one way, < 100 ms RTT | < 1% in any 15 s; burst < 10% in any 200 ms | Client path "including any Wi-Fi and LAN/WAN" |
| Microsoft Teams, customer edge to Microsoft edge | < 15 ms in any 15 s | < 30 ms one way, < 60 ms RTT | < 0.1% in any 15 s; burst < 1% in any 200 ms | ISP segment only |
| Twilio Voice SDK warnings | > 30 ms in 3 of the last 5 samples | RTT > 400 ms, 3 of 5 | > 1%, 3 of 5 | Media at the SDK |
| Google Voice | Receive jitter mean below 30 ms | RTT mean below 100 ms | "As close to zero as possible" | Client to Google media servers |
| Zoom Phone QoS dashboard | Not stated | Not stated | Not stated | MOS only: 3.5 or above good, below 3.5 poor |
Cisco's basis for 30 ms: "extensive lab testing has shown that when jitter consistently exceeds 30 ms voice quality degrades significantly". The rows do not convert into each other: Teams' 30 ms covers one leg, and Y.1541's 50 ms is a percentile spread, not a smoothed mean deviation like RFC 3550 jitter.
Does jitter lower MOS? Only as buffer delay or discarded packets
Jitter lowers MOS only indirectly. The E-model inputs in ITU-T G.107 (06/2015) Table 3 include one-way delay Ta, packet-loss probability Ppl, burst ratio BurstR, codec impairment Ie and robustness Bpl; none is jitter or delay variation.
ITU-T G.114 Annex A.3 describes the conversion: the de-jitter buffer delays packets "to correspond to the delay of the packet with the longest tolerable transit time", and a packet later than that "arrives too late" and is discarded. G.113 (09/2024) adds that "the applicable packet loss is measured at the output of the jitter buffer". Jitter therefore enters as part of Ta (buffer depth) or of Ppl (late discards). For planning, G.114 recommends assuming a buffer adds half its peak delay: 25 ms for a 50 ms delay-variation range.
We computed the tables with the code behind our E-model MOS calculator (method frozen 19 October 2026): narrowband, default R 93.2, echo-free, A = 0, random loss (BurstR = 1), codec values from G.113 Table I.4. Cells are MOS (R), both explained in the MOS score guide; "satisfied" starts at R 80, MOS 4.03 (G.107 Table B.1).
| G.711 with PLC (Ie 0, Bpl 25.1): Ta, ms | 0% | 0.5% | 1% | 2% | 3% | 5% |
|---|---|---|---|---|---|---|
| 50 | 4.41 (93) | 4.37 (91) | 4.33 (90) | 4.23 (86) | 4.13 (83) | 3.92 (77) |
| 100 | 4.41 (93) | 4.37 (91) | 4.33 (90) | 4.23 (86) | 4.13 (83) | 3.92 (77) |
| 150 | 4.41 (93) | 4.37 (91) | 4.32 (89) | 4.23 (86) | 4.13 (83) | 3.92 (77) |
| 200 | 4.34 (90) | 4.30 (88) | 4.24 (87) | 4.14 (83) | 4.02 (80) | 3.79 (74) |
| 250 | 4.18 (84) | 4.11 (82) | 4.05 (81) | 3.92 (77) | 3.78 (74) | 3.53 (69) |
| 300 | 3.96 (78) | 3.89 (77) | 3.81 (75) | 3.66 (71) | 3.52 (68) | 3.24 (63) |
| 400 | 3.56 (69) | 3.47 (67) | 3.38 (65) | 3.21 (62) | 3.05 (59) | 2.75 (53) |
The delay impairment is 0.16 R points at 150 ms and 14.76 at 300 ms. With concealment, G.711 stays "satisfied" at 150 ms up to 3% random loss; without it, or on G.729A, it does not:
| Codec, Ta | 0% | 1% | 2% | 5% |
|---|---|---|---|---|
| G.711, no PLC (Ie 0, Bpl 4.3), 150 ms | 4.41 (93) | 3.83 (75) | 3.25 (63) | 2.16 (42) |
| G.711, no PLC, 300 ms | 3.96 (78) | 3.13 (61) | 2.48 (48) | 1.50 (27) |
| G.729A + VAD (Ie 11, Bpl 19), 150 ms | 4.10 (82) | 3.94 (78) | 3.78 (74) | 3.33 (65) |
| G.729A + VAD, 300 ms | 3.47 (67) | 3.27 (63) | 3.07 (59) | 2.57 (50) |
At 150 ms and 1% loss, G.711 scores 4.32 with Appendix I concealment and 3.83 without it. The 1% limit in the Cisco and Teams documents is generous for the first case and tight for the second.
What is a jitter buffer, and how much delay does it cost?
A jitter buffer is a receive-side queue that holds packets long enough to play them out at the sender's even spacing, "turning variable network delays into constant delays" (Cisco SRND). Too shallow, it discards late packets; too deep, it eats the delay budget.
| Type | How the depth is set | Documented example |
|---|---|---|
| Fixed | Once per call | Asterisk JITTERBUFFER(fixed), max_size default 200 ms; GStreamer rtpjitterbuffer, latency 200 ms |
| Adaptive | Target re-estimated from arrival statistics; playout time-stretched to reach it | libwebrtc NetEq, delay "continuously optimized based on the network conditions"; Asterisk adaptive, target_extra default 40 ms |
A good adaptive buffer grows instantly after an overflow and shrinks slowly (Cisco SRND); resizing causes "infrequent transitional delay variations" (G.114). In browsers, RTCRtpReceiver.jitterBufferTarget (up to 4,000 ms) "influences" the target without setting it (MDN).
The cost of depth, for 80 ms one-way network delay plus 30 ms packetization and codec delay, G.711 with PLC, no loss. Buffer delay is the average the buffer adds, which WebRTC reports as jitterBufferDelay / jitterBufferEmittedCount.
| Average buffer delay | Ta | MOS | R |
|---|---|---|---|
| 20 ms | 130 ms | 4.41 | 93.2 |
| 40 ms | 150 ms | 4.41 | 93.0 |
| 60 ms | 170 ms | 4.39 | 92.4 |
| 100 ms | 210 ms | 4.32 | 89.1 |
| 150 ms | 260 ms | 4.13 | 83.1 |
| 200 ms | 310 ms | 3.92 | 77.4 |
A shallower buffer saves delay but discards more late packets. The late-discard percentages below are illustrative inputs we chose, not measurements.
| Buffer setting | Ta | Loss after buffer | MOS | R |
|---|---|---|---|---|
| 40 ms, 3% late-discard | 150 ms | 3% | 4.13 | 82.9 |
| 60 ms, 1% late-discard | 170 ms | 1% | 4.31 | 88.8 |
| 100 ms, no late-discard | 210 ms | 0% | 4.32 | 89.1 |
| 200 ms, no late-discard | 310 ms | 0% | 3.92 | 77.4 |
The middle wins: 60 ms with 1% discards (4.31) and 100 ms with none (4.32) beat 40 ms (4.13) and 200 ms (3.92). Without concealment the optimum moves deeper: G.711 without PLC scores 4.34 at Ta 200 ms with no loss, and 3.83 at Ta 150 ms with 1% loss.
How to read jitter and jitter-buffer delay from getStats and RTCP
A WebRTC receiver reports these on the audio inbound-rtp stats object:
jitter: the RFC 3550 estimate, in seconds.jitterBufferDelay,jitterBufferTargetDelayandjitterBufferMinimumDelay, each divided byjitterBufferEmittedCount: average buffer delay, average target, and the minimum "purely based on the network characteristics".packetsDiscarded: packets "discarded by the jitter buffer due to late or early-arrival";packetsLost: network loss.concealedSamples÷totalSamplesReceived: share of audio synthesized for lost or late packets.
The counters are cumulative, so take differences over a window; Y.1541 gives 10 to 20 seconds as the minimum interval for telephony:
async function audioWindow(pc, ms = 10000) {
const read = async () => {
for (const s of (await pc.getStats()).values())
if (s.type === "inbound-rtp" && s.kind === "audio") return s;
};
const a = await read();
await new Promise((r) => setTimeout(r, ms));
const b = await read();
const d = (k) => b[k] - a[k];
const out = d("jitterBufferEmittedCount"), rx = d("totalSamplesReceived");
return {
jitterMs: 1000 * b.jitter,
bufferMs: out ? 1000 * d("jitterBufferDelay") / out : null,
lost: d("packetsLost"),
discarded: d("packetsDiscarded"),
concealedPct: rx ? 100 * d("concealedSamples") / rx : null
};
}
For an E-model estimate, Ta ≈ RTT/2 + buffer delay + codec delay, and Ppl = (lost + discarded) ÷ packets expected. On SIP, RTCP receiver reports give jitter (divide by the clock rate) and fraction lost, and the RTCP XR VoIP Metrics block (RFC 3611 §4.7) adds the discard rate and buffer delays in milliseconds. Which counters to collect at each layer is covered in VoIP monitoring.
What causes jitter in a VoIP network, and what reduces it?
Jitter is queueing delay that changes from packet to packet. RFC 3550 says the jitter measure "tracks transient congestion", and G.114 attributes delay variation to packets meeting "varying queue lengths, or different routes through the network".
| Cause | Mechanism | What reduces it |
|---|---|---|
| Voice queued behind bulk traffic | Up to 187.5 ms behind one 1,500-byte packet on a 64 kbps link (Cisco, 2001) | DSCP EF (46) with strict priority (Cisco SRND); EF is "a building block for low delay, low jitter and low loss services" (RFC 3246) |
| Bufferbloat | "Excess buffers in the network cause high latency and latency variation" (RFC 8033) | Active queue management at the bottleneck: PIE (RFC 8033), FQ-CoDel (RFC 8290) |
| Wi-Fi contention | Shared, half-duplex medium; each station waits a random backoff (RFC 8325 §6.1) | WMM, EF mapped to the Voice access category (RFC 8325); 5 GHz, non-overlapping channels (Microsoft) |
| VPN path | Hair-pinning; a farther front door, "introducing extra latency and jitter" (Microsoft) | Split-tunnel VPN for media (Microsoft) |
When the 30 ms rule fails
- Burst loss. G.107 divides Ppl by BurstR, so clustered loss costs more than the same percentage spread out; G.113 calls the independence assumption "unsatisfactory for many real networks". Teams caps burst loss per 200 ms separately. Our tables assume random loss, the optimistic case.
- Averages that hide spikes. Y.1541 bounds the 99.9th percentile because "use of lower quantile values will result in under-estimates of de-jitter buffer size", and wants poor intervals "identified, not obscured". A call average under 30 ms can contain a minute that fails.
- Concealment. A loss threshold without a codec attached is half a threshold, as the G.711 rows with and without PLC show.
- Wideband codecs. The tables are narrowband, and G.113 Appendix I has no Opus values; for an Opus call, use them to compare windows, not to grade the call.
- One direction is not the call. Each direction has its own jitter and buffer.
To see how your stack converts jitter, inject delay variation with tc netem and watch the counters above (WebRTC testing). In production, a scheduled test call records the same counters per window (synthetic monitoring for voice and video); RTCanary's WebRTC room and SIP trunk canaries record RTT, jitter, packet loss and a MOS estimate on every run. Score any measured window in the MOS calculator.
Limitations
- The MOS tables are narrowband E-model planning estimates, not listening tests, with echo-free defaults and A = 0.
- Loss in the tables is random (BurstR = 1); bursty loss of the same size scores lower.
- The trade-off example's late-discard rates and the 80 ms and 30 ms delays are assumptions, not measurements of any path.
- The model treats buffer delay as plain delay; it does not score time-stretching or mid-call resizing.
- G.113 notes that "jitter buffer implementations vary considerably"; the defaults above are documented, not measured.
- Vendor pages change; the thresholds come from the versions in Sources, read in November 2026. Zoom's per-metric "none" covers only the articles we read.
Frequently asked questions
What is an acceptable jitter rate?
Under 30 ms in every vendor target we found: Cisco (average one-way), Teams (client to Microsoft edge, any 15 seconds), Google Voice and Twilio's warning. Teams asks for under 15 ms on the ISP segment alone. Gaming is outside this page.
What is the maximum jitter VoIP can tolerate?
None of the standards we read sets a hard maximum for RTP jitter. The practical limit is what your buffer absorbs without pushing one-way delay past 150 ms or discarding more than the codec conceals; Y.1541 plans for 50 ms of delay variation at the 99.9th percentile.
What is the maximum one-way latency that VoIP can support?
ITU-T G.114 sets 400 ms one way as the limit for general network planning, with exceptions such as an unavoidable double satellite hop, and expects "essentially transparent interactivity" below 150 ms. In the E-model, G.711 with PLC at 400 ms scores 3.56.
What is latency vs jitter?
Latency is the time a packet takes; jitter is how much that time varies between packets. The jitter buffer turns jitter into latency, since every millisecond of buffering adds to mouth-to-ear delay.
What is acceptable packet loss for VoIP?
Cisco targets 1% or less, Teams under 1% in any 15 seconds with bursts under 10% in any 200 ms, Y.1541 0.1% on the network, and G.1010 under 3% assuming concealment. Count late discards too: G.113 measures loss after the jitter buffer.
What is acceptable jitter for video conferencing?
Cisco's SRND gives interactive video the voice targets (jitter 30 ms, one-way latency 150 ms, loss 1%) because videoconferencing "includes a G.711 audio codec for voice". RFC 3550 notes that packets of one video frame share a timestamp but leave at different times, which inflates computed video jitter.
Sources
- ITU-T G.114 (05/2003), One-way transmission time: clause 4, Annex A.3.
- ITU-T G.107 (06/2015), The E-model: Tables 3 and B.1; ITU-T G.109 (09/99).
- ITU-T G.113 (09/2024): Appendix I; ITU-T G.1010 (11/2001): clause 4.2, Table I.1.
- ITU-T Y.1541 (12/2011): Tables 1 and 2, clause 5.3.
- RFC 3550, RTP: 6.4.1, 6.4.4, A.8; RFC 3551; RFC 7587; RFC 3611: 4.7.
- RFC 3246; RFC 8033; RFC 8290; RFC 8325.
- W3C, Identifiers for WebRTC's Statistics API, Candidate Recommendation Draft, 25 September 2025; MDN, jitterBufferTarget.
- libwebrtc NetEq; Asterisk JITTERBUFFER; GStreamer rtpjitterbuffer.
- Cisco, Enterprise QoS Solution Reference Network Design Guide 3.3 (2005): chapter 1; Cisco, Quality of Service for Voice over IP (2001).
- Microsoft Learn, Understand Teams network requirements; Prepare your organization's network for Teams.
- Twilio, Voice Insights SDK call quality events; Google, Troubleshoot Google Voice call quality; Zoom, Viewing the Zoom Phone quality of service dashboard.