Viewers notice sound that arrives just 45 ms before the picture, but tolerate up to 125 ms of sound arriving after it, the asymmetric detectability window behind the international broadcast recommendation on sound and vision timing (ITU-R BT.1359-1, 1998). Real playback chains often miss that window: a November 2026 photodiode study measured median sound lags of 79.2 to 210.9 ms across 24 player and resolution combinations, with up to 88.4% of trials beyond the detectability threshold (Ho and Li, Preprints.org 2026). Tolerance also depends on what is on screen, with acceptability thresholds ranging from 200 ms for a newsreader to 1.3 s for American football (Ebenezer, arXiv 2022). Lip-sync drift (the gap between when a mouth moves and when its sound is heard) is a different problem from caption timing, which we covered in our closed caption delay and accuracy statistics. We aggregated data from the EBU, the ATSC, SMPTE, IEEE Transactions on Broadcasting, the Journal of the Acoustical Society of America, PLOS ONE, Allion Labs, Amazon Science and a dozen other primary sources.
TL;DR
- Average detectability threshold: about 45 ms of audio lead to 125 ms of audio lag (ITU-R BT.1359-1, 1998).
- European broadcast limit: 40 ms of sound before picture and 60 ms after, end to end (EBU R37, 2007).
- US DTV guidance at the encoder input: no more than 15 ms of audio lead or 45 ms of lag (ATSC IS/191, reported by TV Technology 2004).
- Mean audio-leading detection threshold in an Intel lab test: 185.19 ms, SD 42.32 ms (Younkin and Corriveau, IEEE Transactions on Broadcasting 2008).
- Speech judged synchronous from 131 ms of audio lead to 225 ms of visual lead (Conrey and Pisoni, JASA 2006).
- Acceptability threshold for audio leading video: 200 ms for a newsreader vs 1.3 s for American football (Ebenezer, arXiv 2022).
- Median delivered offsets in tested talking-head playback: sound late by 79.2 to 210.9 ms in every cell (Ho and Li, Preprints.org 2026).
- Zoom measured from 92 ms of video lead to 73 ms of audio lead across five computers (Allion Labs AV-Sync testing).
- Worst average audio-visual onset gap on an uncalibrated mixed reality headset: 156.63 ms (Eckhoff et al., PLOS ONE 2024).
- Synchrony window for a native English talker: 100 ms in young normal-hearing listeners vs 250 ms in older listeners (Gordon-Salant et al., Frontiers in Psychology 2022).
- Lip-sync classifier accuracy on the LRS2 dataset: 96.5% for IC-SyncNet vs 75.8% for the original SyncNet (Park et al., arXiv 2024).
1. Detection Thresholds: When Viewers Notice Lip-Sync Error
The single most consistent finding across 80 years of research is asymmetry. Sound that arrives early is caught at roughly a third of the offset that late sound needs: about 45 ms of audio lead against 125 ms of audio lag in the subjective tests (tests in which viewers rate clips) summarized by the ITU-R BT.1359-1 recommendation. Nature explains the bias: light travels faster than sound, and in natural speech the mouth starts moving before the voice. In the speech stimulus analyzed by Venezia et al., mouth closure preceded the sound by 163 ms, so the brain is already trained to accept late audio.
The spread between studies matters as much as the averages. A 1940s Bell Laboratories study put detection at 35 ms of lead and 100 ms of lag, while an Intel lab test of audio-leading errors found a much looser mean of 185.19 ms (Younkin and Corriveau, IEEE Transactions on Broadcasting 2008). The practical reading: thresholds are averages over a population, and the most sensitive viewers sit well inside them. A stream engineered to the average still annoys a share of its audience.
| Metric | Value | Source |
|---|---|---|
| Average detectability threshold, sound leading vs sound lagging the picture | About +45 ms to -125 ms | ITU-R BT.1359-1 (1998) |
| Detection threshold in an early Bell Laboratories study, audio leading vs audio lagging | More than 35 ms lead or more than 100 ms lag | Bell Laboratories study, reported in TV Technology (Hoffner) 2003 |
| Mean detection threshold for audio leading the lips | 185.19 ms (SD 42.32 ms) | Younkin and Corriveau, IEEE Transactions on Broadcasting 2008 |
| Speech judged synchronous, largest audio lead vs largest visual lead (30 adults) | 131 ms audio lead to 225 ms visual lead | Conrey and Pisoni, JASA 2006 |
| Non-speech (flash and tone) judged synchronous, audio lead vs visual lead | 153 ms audio lead to 247 ms visual lead | Conrey and Pisoni, JASA 2006 |
| Width of the synchrony window at half its peak, speech vs non-speech | 357 ms vs 400 ms | Conrey and Pisoni, JASA 2006 |
| Detection threshold for speech, video leading sound vs sound leading video | 258 ms vs 131 ms | Dixon and Spitz 1980, cited in Venezia et al., Attention, Perception, and Psychophysics 2016 |
| Detection threshold for a hammer hitting a peg, video leading vs sound leading | 188 ms vs 75 ms | Dixon and Spitz 1980, cited in Venezia et al. 2016 |
Context note: Conrey and Pisoni found the center of the synchrony window sat at 47 ms of visual lead, not at zero, for both speech and non-speech (Conrey and Pisoni, JASA 2006). The Venezia et al. review traces the same pattern across replications. Most recent available data for population thresholds: Younkin and Corriveau 2008; no newer large-sample threshold study for broadcast lip-sync surfaced in our research.
2. Broadcast Tolerances and Sync Budgets
Perception studies tell you where viewers notice; standards tell engineers how much of that margin each part of the chain may spend. The striking pattern is how much tighter the engineering limits are than the perception limits. EBU R37 allows each stage only 5 ms early and 15 ms late, while the whole chain must stay inside a 40 ms lead and 60 ms lag end-to-end limit (EBU R37, 2007). EBU states its limits are based on the delay at which 50% of observers notice lip-sync failure.
The US went tighter still. ATSC’s Implementation Subcommittee found the international 1998 limits inadequate for digital television and asked that audio entering the encoder never lead video by more than 15 ms or lag by more than 45 ms (TV Technology, ATSC Recommends AV Sync Rules, 2004). For comparison, the old NTSC rule of thumb was +1 and -2 frames, or +33 and -66 ms. The 1998 recommendation also splits its own budget, leaving segments outside the broadcaster’s control only ±2 ms.
| Metric | Value | Source |
|---|---|---|
| Overall sound-to-picture tolerance (also the average acceptability threshold) | +90 ms to -185 ms | ITU-R BT.1359-1 (1998) |
| Budget split: camera source to zero reference point (the studio timing reference) vs zero reference to transmitter input | +25 / -100 ms vs +22.5 / -30 ms | ITU-R BT.1359-1 (1998) |
| End-to-end limit at any output intended for emission, sound before vs sound after picture | 40 ms vs 60 ms | EBU R37 (2007) |
| Limit per individual stage of the production chain | 5 ms early to 15 ms late | EBU R37 (2007) |
| DTV limit for audio at the encoder input, lead vs lag | 15 ms vs 45 ms | ATSC IS/191, reported by TV Technology 2004 |
| DTV limit for the whole chain, audio leading vs audio lagging video | +30 ms to -90 ms | ATSC IS/191, reported by TV Technology 2004 |
| IPTV quality-of-experience limit, audio leading vs audio lagging video (provisional, standard definition) | 15 ms vs 45 ms maximum | ITU-T G.1080 (2008) |
| Analog NTSC rule of thumb in frames and milliseconds | +1 / -2 frames (+33 / -66 ms) | TV Technology (Hoffner) 2004 |
Outlier note: these documents are old but still current. The sound and vision recommendation was approved on November 30, 1998 and the IPTV quality recommendation on December 7, 2008, and both are listed as in force. Most recent available data: EBU R37, 2007.
3. Content Type and Audience: Who Tolerates How Much
A talking face is the hardest test a stream can face. In a 20-viewer study of 45 clips, the acceptability threshold for audio leading video was 200 ms for a newsreader but 1.3 s for American football (Ebenezer, arXiv 2022). Wide sports shots offer few sharp audio-visual events to anchor timing, so a gaming or sports stream with the streamer’s face cam on screen is judged by the stricter, face-based standard.
The audience matters too. With a native English talker, older listeners show synchrony windows (the range of offsets still perceived as simultaneous) at least twice as wide as young adults, yet still lose speech recognition when the signal is misaligned, even when they do not notice the error (Gordon-Salant et al., Frontiers in Psychology 2022). A wide window is not the same as immunity: sync errors can cost comprehension without ever being reported as a complaint.
| Metric | Value | Source |
|---|---|---|
| Acceptability threshold for audio leading video, newsreader clip | 200 ms | Ebenezer, arXiv 2022 |
| Acceptability threshold for audio leading video, cricket clip | 500 ms | Ebenezer, arXiv 2022 |
| Acceptability threshold for audio leading video, soccer and tennis clips | 1 s | Ebenezer, arXiv 2022 |
| Acceptability threshold for audio leading video, American football clip | 1.3 s | Ebenezer, arXiv 2022 |
| Synchrony window for a native English talker, young normal-hearing vs older listeners (17 per group) | 100 ms vs 250 ms | Gordon-Salant et al., Frontiers in Psychology 2022 |
| Typical synchrony window width, visual-leading pairs vs audio-leading pairs | 200-250 ms vs 100-150 ms | Cecere et al., European Journal of Neuroscience 2016 |
| Baseline sensitivity score (d prime, higher means better detection), audio-leading vs visual-leading pairs, 30 adults | 1.58 vs 0.98 | Cecere et al., European Journal of Neuroscience 2016 |
| Participants in a 2026 speech-in-noise asynchrony study, cochlear implant users vs typically hearing adults (offsets up to 400 ms) | 15 vs 31 | Salagovic et al., Journal of Speech, Language, and Hearing Research 2026 |
Context note: the Cecere et al. sensitivity gap was statistically robust (F(1,27) = 37.10, p < 0.0001) (Cecere et al., European Journal of Neuroscience 2016). In the 2026 study, intelligibility peaked near synchrony for both groups, and cochlear implant users showed larger drops as asynchrony grew (Salagovic et al., JSLHR 2026). The sports study is a preprint and is older than three years; most recent available data: Ebenezer, 2022.
4. Measured Offsets in Players, Calling Apps, and Headsets
Standards assume each device reports and corrects its own delay. Measurements show that assumption failing. A November 2026 photodiode and line-output study of four talking-head video systems, three players and two resolutions found sound lagging in every one of 24 cells, with median offsets of 79.2 to 210.9 ms, and 60.9% to 88.4% of 3,637 trials beyond the broadcast detectability threshold (Ho and Li, Preprints.org 2026). Their own measurement uncertainty between channels was 0.0144 ms, so the error is in the playback chain, not the instrument.
Calling and streaming apps scatter on both sides of zero. In Allion Labs testing with one USB webcam on five computers, the same app could lead on one machine and lag on another (Allion Labs AV-Sync testing). That is why packet-level fixes, covered in our WebRTC packet loss concealment statistics, cannot by themselves guarantee lip-sync: sync is decided at the endpoints. Every processing stage a streamer adds to the microphone path, from noise suppression to a voice changer in OBS, adds its own delay to that budget, which is why EBU R37 budgets each stage separately.
| Metric | Value | Source |
|---|---|---|
| Median delivered audio-video offset across 24 player and resolution cells (negative means sound late) | -210.9 ms to -79.2 ms | Ho and Li, Preprints.org 2026 |
| Share of trials beyond the broadcast detectability threshold, range across cells | 60.9% to 88.4% | Ho and Li, Preprints.org 2026 |
| Share of trials beyond the broadcast acceptability threshold, range across cells | 31.6% to 59.4% | Ho and Li, Preprints.org 2026 |
| Zoom, range across five computers (positive means audio first) | -92 ms to +73 ms | Allion Labs AV-Sync testing (undated) |
| Microsoft Teams, range across five computers | -95 ms to -20 ms | Allion Labs AV-Sync testing (undated) |
| Twitch, range across four computers | -90 ms to +32 ms | Allion Labs AV-Sync testing (undated) |
| Google Meet, range across five computers | -100 ms to +42 ms | Allion Labs AV-Sync testing (undated) |
| Largest average audio-visual onset gap on an uncalibrated mixed reality headset, with jitter up to | 156.63 ms, jitter up to ±11.82 ms | Eckhoff et al., PLOS ONE 2024 |
Context note: calibration works. Delaying the picture on a Varjo XR-3 moved its gap from 53.49 ± 3.15 ms to -3.89 ± 1.56 ms over 100 presentations (Eckhoff et al., PLOS ONE 2024). The photodiode study at the top of this section was posted on November 3, 2026 and has not yet been peer reviewed.
5. Long-Session Drift: Clocks, Capture Cards, and Sync Measurement
Static offsets can be dialed out once. Drift cannot, because it grows with time. When audio and video are clocked independently, a frequency error accumulates for as long as the stream runs. A clock at the ±1 ppm limit of a Grade 1 digital audio reference (DARS, the master clock that sets the audio sample rate) drifts 3.6 ms per hour: 0.000001 x 3,600 seconds = 0.0036 s, or 28.8 ms over an eight-hour stream (our calculation from the Grade 1 tolerance in the table below). Broadcast plants avoid this by locking everything to one reference, the same logic behind the IP migration tracked in our AoIP, Dante and AES67 broadcast statistics.
Consumer capture hardware is looser. Software vendor LEADTOOLS documents capture devices that start 60 to 90 ms out of sync and devices that begin drifting after 8 to 24 hours (LEADTOOLS, Timestamp Drifting in Live Capture Situations). For a 24/7 stream, drift is not an edge case; it is a scheduled maintenance item.
| Metric | Value | Source |
|---|---|---|
| Long-term frequency accuracy of a Grade 1 digital audio reference | Within ±1 ppm | ITU-R BS.2032-0 (2013) |
| Drift at the ±1 ppm limit, per hour and over 8 hours (our calculation) | 3.6 ms per hour; 28.8 ms per 8 hours | Calculated from ITU-R BS.2032-0 (2013) |
| Capture cards observed drifting after continuous capture of | 8 to 24 hours | LEADTOOLS documentation (vendor’s own observation) |
| Initial audio-video offset on some capture devices | About 60-90 ms | LEADTOOLS documentation (vendor’s own observation) |
| Timestamp difference on capture devices with mismatched clocks | Over 500 ms | LEADTOOLS documentation (vendor’s own observation) |
| Accuracy of the BBC open-source sync measurement rig (light sensor plus audio input) | Approximately millisecond accuracy | BBC R&D dvbcss-synctiming |
| Example pass/fail tolerance in the BBC rig, companion screen app vs TV | ±8 ms vs ±23 ms | BBC R&D dvbcss-synctiming |
| Transport methods for lip-sync fingerprints (SDI, UDP over IP, MPEG-2 transport stream), one container per frame | 3 | SMPTE ST 2064-2 (2015) |
Context note: the BBC rig plays test video with flashes and beeps and times them at the screen and speaker, the only point where sync is actually experienced (BBC R&D dvbcss-synctiming). SMPTE’s fingerprint approach measures timing changes without test signals by comparing fingerprints taken upstream and downstream (SMPTE ST 2064-2:2015).
6. Automated Lip-Sync Detection: What Machines Can and Cannot Catch
Streaming services now check sync with machine learning rather than human spot checks. Prime Video built detectors for 18 types of defect, including audio-video sync, and runs its lip-sync detector on four-second clips centered on a single face across thousands of channels and live events (Amazon Science, How Prime Video uses machine learning to ensure video quality, 2022). Classifier accuracy has climbed fast: 96.5% on the LRS2 dataset in 2024, against 75.8% for the original SyncNet (Park et al., Interpretable Convolutional SyncNet, arXiv 2024).
The catch is that file-level scores do not see the playback chain. Ho and Li found SyncNet’s sync scores uncorrelated with the offsets they measured at the screen, and SyncNet’s own offset estimate agreed with the measurement only within a 282.2 ms half-width. A clean file can still play out of sync, and the second check has to happen at the device.
| Metric | Value | Source |
|---|---|---|
| Video defect types covered by Prime Video detectors (audio-video sync among them) | 18 | Amazon Science (Prime Video) 2022 |
| Tennis audio-video sync error database built by Amazon researchers | 504,300 frames from 6 hours of video | Ebenezer et al., ICASSP 2021 (Amazon) |
| Sync classification accuracy, IC-SyncNet on LRS2 vs LRS3 | 96.5% vs 93.8% | Park et al., arXiv 2024 |
| Sync classification accuracy, original SyncNet on LRS2 | 75.8% | Park et al., arXiv 2024 |
| Human ratings collected for a perceptual sync quality dataset (100+ hours, 80+ annotators) | 120,000 | Goncalves et al. (Amazon), PEAVS, ECCV 2024 |
| Correlation of the PEAVS metric with human ratings, set level vs clip level | 0.79 vs 0.54 | Goncalves et al. (Amazon), PEAVS, ECCV 2024 |
| Correlation between SyncNet sync scores and measured offset magnitude | r = -0.066 to +0.083 | Ho and Li, Preprints.org 2026 |
Context note: in the PEAVS data, both human scores and the metric peaked above 4 on a 5-point scale at a 45 ms audio shift and fell as the shift grew (Goncalves et al., PEAVS, ECCV 2024). Amazon’s tennis method flags an error when a racquet hit appears in the audio but not in nearby video (Amazon Science, Detection of audio-video synchronization errors via event detection).
Summary: Live Stream Audio-Video Sync (Lip-Sync) Drift by the Numbers
| Metric | Value | Source |
|---|---|---|
| Average detectability threshold, sound leading vs lagging | About +45 ms to -125 ms | ITU-R BT.1359-1 (1998) |
| Width of the synchrony window at half its peak, speech vs non-speech | 357 ms vs 400 ms | Conrey and Pisoni, JASA 2006 |
| End-to-end emission limit, sound before vs after picture | 40 ms vs 60 ms | EBU R37 (2007) |
| Per-stage limit in the production chain | 5 ms early to 15 ms late | EBU R37 (2007) |
| DTV limit at the encoder input, lead vs lag | 15 ms vs 45 ms | ATSC IS/191, via TV Technology 2004 |
| DTV whole-chain limit | +30 ms to -90 ms | ATSC IS/191, via TV Technology 2004 |
| Mean detection threshold for audio leading the lips | 185.19 ms (SD 42.32 ms) | Younkin and Corriveau, IEEE Transactions on Broadcasting 2008 |
| Speech judged synchronous, audio lead vs visual lead | 131 ms vs 225 ms | Conrey and Pisoni, JASA 2006 |
| Acceptability threshold for audio lead, newsreader vs American football | 200 ms vs 1.3 s | Ebenezer, arXiv 2022 |
| Synchrony window, young vs older listeners | 100 ms vs 250 ms | Gordon-Salant et al., Frontiers in Psychology 2022 |
| Sensitivity, audio-leading vs visual-leading pairs | d prime 1.58 vs 0.98 | Cecere et al., European Journal of Neuroscience 2016 |
| Median delivered offsets in tested talking-head playback | Sound late by 79.2 to 210.9 ms | Ho and Li, Preprints.org 2026 |
| Microsoft Teams offset range across five computers | -95 ms to -20 ms | Allion Labs AV-Sync testing |
| Zoom offset range across five computers | -92 ms to +73 ms | Allion Labs AV-Sync testing |
| Worst average onset gap, uncalibrated mixed reality headset | 156.63 ms | Eckhoff et al., PLOS ONE 2024 |
| Initial offset on some capture devices | About 60-90 ms | LEADTOOLS documentation (vendor) |
| Lip-sync classifier accuracy on LRS2, IC-SyncNet vs SyncNet | 96.5% vs 75.8% | Park et al., arXiv 2024 |
| PEAVS correlation with human ratings, set level | 0.79 | Goncalves et al. (Amazon), ECCV 2024 |
Methodology and Sources
- ITU recommendations: ITU-R BT.1359-1, Relative timing of sound and vision for broadcasting (1998); ITU-T G.1080, Quality of experience requirements for IPTV services (2008); ITU-R BS.2032-0, Synchronization of digital audio sample clock to video references (2013).
- EBU R37, Relative timing of the sound and vision components of a television signal (2007).
- ATSC IS/191 and the Bell Laboratories threshold, via Randy Hoffner’s columns in TV Technology: ATSC Recommends AV Sync Rules (2004) and A/V Synchronization: How Bad Is Bad? (2003).
- Younkin and Corriveau, Determining the Amount of Audio-Video Synchronization Errors Perceptible to the Average End-User, IEEE Transactions on Broadcasting (2008).
- Conrey and Pisoni, Auditory-visual speech perception and synchrony detection for speech and nonspeech signals, JASA (2006).
- Venezia et al., Timing in Audiovisual Speech Perception: A Mini Review and New Psychophysical Data, Attention, Perception, and Psychophysics (2016).
- Cecere, Gross and Thut, European Journal of Neuroscience (2016).
- Gordon-Salant et al., Detection and Recognition of Asynchronous Auditory/Visual Speech, Frontiers in Psychology (2022).
- Salagovic, Stevenson and Butler, Journal of Speech, Language, and Hearing Research (2026).
- Ebenezer, A subjective study of the perceptual acceptability of audio-video desynchronization in sports videos, arXiv (2022).
- Ho and Li, Software Lip-Sync Metrics Do Not Predict Delivered Audio-Visual Synchronisation, Preprints.org (2026).
- Allion Labs, Audio-to-Video Synchronization (AV-Sync) Testing of Webcam and Video Software.
- Eckhoff, Schnupp and Cassinelli, Temporal precision and accuracy of audio-visual stimuli in mixed reality systems, PLOS ONE (2024).
- LEADTOOLS, Timestamp Drifting in Live Capture Situations.
- BBC R&D, dvbcss-synctiming and SMPTE ST 2064-2:2015, Audio to Video Synchronization Measurement, Fingerprint Transport.
- Amazon: Prime Video video quality analysis (2022), Detection of audio-video synchronization errors via event detection, ICASSP (2021) and Goncalves et al., PEAVS, ECCV (2024).
- Park et al., Interpretable Convolutional SyncNet, arXiv (2024).
- Data watch: the detectability figures conflict across studies. The 1998 broadcast recommendation puts audio-lead detection at about 45 ms, the Bell Laboratories study at 35 ms, and Younkin and Corriveau at a mean of 185.19 ms; we report all three rather than pick one, and differences in stimuli and method likely explain the gap. Sign conventions also differ: most broadcast documents and Allion Labs use positive for audio first, while the 2026 photodiode study reports sound lag as negative. The Allion Labs page carries no publication date, and the photodiode study (posted November 3, 2026) and the Ebenezer sports study are preprints that have not been peer reviewed. LEADTOOLS figures are a software vendor’s own observations. The ATSC IS/191 values come from TV Technology’s reporting because the ATSC document itself was not reachable. The 3.6 ms per hour drift figure is our calculation from the ±1 ppm tolerance, not a measured value. Most of the threshold and tolerance data is older than three years and flagged as most recent available.
Last updated: November 6, 2026. We update this roundup quarterly, and the next refresh is expected when the peer-reviewed version of the November 2026 photodiode study on delivered lip-sync is published.