Hearing your own voice just 25 to 30 milliseconds late is already rated ‘disturbing’ by people reading aloud, according to a 32-participant study in Ear and Hearing, and that number goes a long way to explain why presenter earpiece feeds in live television are built around removing the presenter’s own voice. The pictures are judged just as harshly: the EBU allows sound to run no more than 40 ms ahead of the picture or 60 ms behind it at any output intended for emission. Meanwhile the production itself is moving away from the venue, with two-thirds of booked remote services at the Paris 2024 Olympics carried over a cloud platform rather than satellite. We aggregated data from the ITU, the EBU, ATSC, hearing research published in Ear and Hearing and PLoS ONE, a BBC R&D, University of Strathclyde and TV2 field study, the AIMS alliance, IFIP Networking satellite measurements, and Haivision’s 2026 industry survey. For the on-air side of live commentary, see our companion roundup on esports caster and commentary broadcast audio.
Two terms run through this article. Talkback (also called intercom or comms) is the private voice channel between the control room, which British broadcasters call the gallery, and the crew. IFB, short for interruptible foldback, is the feed in a presenter’s earpiece: normally the programme sound, which a producer can interrupt to give instructions. That feed is usually a mix-minus, meaning the full mix minus the presenter’s own voice, so they never hear themselves come back late.
TL;DR
- Viewers start to detect a sound/picture timing error at about +45 ms (sound early) to -125 ms (sound late) (ITU-R BT.1359-1, 1998).
- At any output intended for emission, sound may lead the picture by 40 ms or less and lag by 60 ms or less (EBU R37-2007).
- At the input to DTV encoders, sound should never lead by more than 15 ms or lag by more than 45 ms (ATSC IS-191).
- Reading aloud, 32 participants rated a 25 to 30 msec delay of their own voice as disturbing (Stone and Moore, Ear and Hearing 2002).
- The threshold for noticing that your own voice is delayed sat at approximately 100 ms, with an SD of 66.54 ms (Yamamoto and Kawabata, PLoS ONE 2011).
- Current DECT talkback systems reach below 40 ms to a wireless beltpack and below 70 ms beltpack to beltpack (BBC R&D, Strathclyde and TV2, IBC2024).
- SMPTE ST 2110-30 Levels B and C use 125-microsecond audio packets, against 1 ms in Level A (AIMS, AES67 and ST 2110-30 introduction).
- Starlink’s median TLS handshake latency was approximately 80 ms in 2024 (Richter, Ververis and Bajpai, IFIP Networking 2025).
- 41% of more than 1,300 surveyed professionals called remote production their biggest technology priority for 2026 (Haivision Broadcast Transformation Report 2026).
- Two-thirds of booked remote services across 54 broadcasters at Paris 2024 used OBS Live Cloud (Olympic Broadcasting Services, 2024).
1. Lip-Sync Tolerances: What Broadcast Standards Allow
Every talkback and IFB decision sits inside a picture-timing budget that viewers police without knowing it. The asymmetry is the key insight: viewers tolerate sound arriving late (up to about -125 ms before they detect it) far better than sound arriving early (about +45 ms), a pattern that matches everyday experience, where the sight of an event reaches us before its sound. In these standards, a positive number means the sound is ahead of the picture and a negative number means it is behind.
The standards then split that perceptual budget across the chain. The EBU figures were set from subjective tests of when a broken match between lip movements and speech becomes perceptible to 50% of observers, and the EBU asks for only 5 ms early to 15 ms late at each individual stage, so that errors from many stages do not pile up beyond the end-to-end limit. ATSC went tighter still at the encoder input, because a second half of the chain, from encoder to receiver, still has to fit inside the viewer’s tolerance.
| Metric | Value | Source |
|---|---|---|
| Detectability threshold for sound/picture timing error (positive = sound ahead of picture) | About +45 ms to -125 ms | ITU-R BT.1359-1 (1998) |
| Acceptability threshold for sound/picture timing error | About +90 ms to -185 ms | ITU-R BT.1359-1 (1998) |
| Allowed timing error from the final programme source selection point (for example master control) to the transmitter input | +22.5 ms to -30 ms | ITU-R BT.1359-1 (1998) |
| End-to-end limit at any output intended for emission: sound before picture vs sound after picture | 40 ms or less vs 60 ms or less | EBU R37-2007 |
| Required accuracy at each stage of the programme chain | Sound 5 ms early to 15 ms late | EBU R37-2007 |
| Sound timing at the inputs to DTV encoding devices | Never more than 15 ms early or 45 ms late | ATSC IS-191, as reported by TV Technology, January 7, 2004 |
| Overall end-to-end range for the DTV system | +30 ms to -90 ms | ATSC IS-191, as reported by TV Technology, 2004 |
Context note: these documents are old but still the working references. Most recent available data: Recommendation BT.1359-1 (1998), R37 (first issued 1986, re-issued 2007) and ATSC IS-191 (reported 2004). The R37 text notes that particular care is needed in HD production, so read the limits as ceilings, not targets.
2. Hearing Your Own Voice Late: Self-Monitoring Delay Thresholds
This is the number that shapes IFB design. Hearing-aid research is the best controlled evidence on how a talker reacts to their own delayed voice, and it converges on a narrow window: delays become disturbing somewhere between 20 and 30 msec, long before speech itself breaks down. For a presenter, that means a mix-minus is not a nicety. Any path that loops the presenter’s voice back through a codec, a network and a return feed will cross that threshold many times over (compare the one-way satellite figures in section 5).
The second insight is that noticing a delay and being bothered by it are different things. In a small headphone experiment, people only reliably sensed that their own voice was late at around 100 ms, yet the disturbance studies place discomfort at a fraction of that. Adaptation helps a little, but only after minutes to an hour, which guests on a live show never get. Producers who handle in-studio monitoring will recognise the same arithmetic from computer audio buffers, covered in our roundup on DAW buffer size and audio latency.
| Metric | Value | Source |
|---|---|---|
| Delay at which the echo of one’s own voice was rated disturbing (20 normally hearing listeners with simulated mild to moderate hearing loss) | Between 20 and 30 msec | Stone and Moore, Ear and Hearing 1999 |
| Reading aloud through a real-time digital hearing aid, 32 participants (16 men, 16 women): delay rated disturbing vs delay at which speech production measures changed | 25 to 30 msec vs above 30 msec | Stone and Moore, Ear and Hearing 2002 |
| 25 hearing-impaired talkers: delay range given a rating of 3 (probably just acceptable), and time over which some acclimatization occurred | 14 to 30 msec; about 1 hour | Stone and Moore, Ear and Hearing 2005 |
| Point at which a delayed own voice felt simultaneous, with delays of 0 to 200 ms through headphones (6 adults aged 21 to 24 repeating a single voice sound) | 98.41 ms (SD 66.54 ms); detection threshold approximately 100 ms | Yamamoto and Kawabata, PLoS ONE 2011 |
| Same study: point of felt simultaneity after 3 minutes of adaptation to 0, 66 or 133 ms of delay | 67.60, 92.23 and 112.10 ms (F(2,10) = 15.80, p<.001) | Yamamoto and Kawabata, PLoS ONE 2011 |
| Processing delay tested on 20 normally hearing and 20 hearing-impaired participants rating annoyance on a 7-point scale | Up to 50 ms; tolerance significantly greater with hearing loss in two of three voice conditions | Goehring, Chapman, Bleeck and Monaghan, International Journal of Audiology 2018 |
Context note: the samples are small and specific (normally hearing adults, hearing-impaired talkers, six young adults aged 21 to 24), the delays were created by hearing-aid processors or headphones, and none of the participants were broadcast presenters. Most recent available data: Stone and Moore, 1999 to 2005; Yamamoto and Kawabata, 2011; Goehring et al., 2018.
3. Talkback, IFB and Intercom Latency Budgets
Different people on a live production live inside very different latency budgets. A performer’s in-ear monitor needs about 4 ms, a crew member on a wireless talkback beltpack (the belt-worn intercom unit) can work with tens of milliseconds, and a remote vision mixer operator (the person cutting between cameras) can tolerate more than half a second as long as nobody at the venue needs to see the result in real time. The budgets differ because the feedback loops differ: a singer hears their own voice, a camera operator only hears instructions.
That is also why IFB is engineered as a mix-minus rather than a full programme return. A 2004 international report on contribution loops lists using a mix-minus or clean feed to foldback, muting foldback when microphones are open, and closed headphones or in-ear monitoring as the operating practices that keep echo out of the loop (ITU-R Report BT.2044). For comparison with another voice-critical network, push-to-talk radio for first responders runs on similar budgets, covered in our roundup on mission-critical push-to-talk latency.
| Metric | Value | Source |
|---|---|---|
| Latency target for performers’ in-ear monitors vs one-way latency of radio microphones and in-ear monitors in current use | 4 ms vs under 3 ms | BBC R&D, University of Strathclyde and TV2, IBC2024 |
| Digital (DECT) production talkback: base to wireless beltpack vs beltpack to beltpack | Below 40 ms vs below 70 ms | BBC R&D, University of Strathclyde and TV2, IBC2024 |
| Clear-Com FreeSpeak Edge 5 GHz beltpack: audio latency vs beltpack to beltpack | Under 35 ms vs 60 ms (manufacturer’s own claim) | Clear-Com FSE-BP50 datasheet, revision August 2023 |
| Mouth-to-ear delay below which most speech and non-speech applications see essentially transparent interactivity (400 ms is the general network planning ceiling) | 150 ms | ITU-T G.114 (2003) |
| Operator input-to-response delay: point where accuracy and speed drop vs point of significantly reduced sense of control | About 100 ms vs beyond 300 ms | BBC R&D, University of Strathclyde and TV2, IBC2024 |
| Camera-to-vision-mixer delay that remote vision mixer operators can tolerate when no return feeds to the venue are needed | Over 600 ms | BBC R&D, University of Strathclyde and TV2, IBC2024 |
Outlier note: the 150 ms telephony figure is a conversation benchmark, not an IFB benchmark. A presenter hearing their own voice reacts at 20 to 30 msec (section 2), while two people talking over a link tolerate roughly five times more (150 ms against 30 msec).
4. AoIP Transport: AES67 and SMPTE ST 2110-30 Timing
Inside the building, the transport itself has stopped being the bottleneck. Audio over IP (AoIP) standards define packet times in milliseconds or less, so the choice between a 1 ms and a 125-microsecond packet time is now a configuration decision, not a limit of the technology. The real latency sits elsewhere: in codecs, buffers and long-distance links (section 5).
The second story is interoperability. Proprietary AoIP systems each used their own discovery, clocking and stream formats, so a talkback matrix from one vendor could not simply subscribe to a console from another. AES67 defines a common operating mode, and SMPTE ST 2110-30 constrains it further for broadcast, which lets IP intercom panels and beltpack base stations share a network with cameras and consoles. Our roundup on AoIP, Dante and AES67 studio migration covers how far that migration has progressed.
| Metric | Value | Source |
|---|---|---|
| AES67 first published, and later updates | 2013; updated 2015 and 2018 | AIMS, Introduction to AES67 and its relationship with SMPTE ST 2110-30 |
| SMPTE ST 2110 suite: initial sections published | December 2017 | AIMS, AES67 and ST 2110-30 introduction |
| ST 2110-30 primary operating point (Level A): packet time and channels per stream | 1 ms; 2 and 8 channels | AIMS, AES67 and ST 2110-30 introduction |
| ST 2110-30 Levels B and C: packet time and maximum channels per stream | 125 microseconds; up to 8 (Level B) and up to 64 (Level C) | AIMS, AES67 and ST 2110-30 introduction |
| AES67 compatibility mode: common audio format | 24-bit linear (L24), up to 8 channels per stream, 48 kHz | AIMS, AES67 and ST 2110-30 introduction |
| Native stream limits before AES67 mode: Dante vs RAVENNA | Up to 4 channels per flow vs up to 64 channels per stream | AIMS, AES67 and ST 2110-30 introduction |
Context note: AIMS (the Alliance for IP Media Solutions) is an industry alliance whose roadmap is built on these standards. The figures above are design parameters from the standards, not field measurements of any particular installation.
5. Contribution Links: Satellite, Private 5G, Cellular and LEO
Once the signal leaves the building, physics and codecs take over. A single geostationary satellite hop adds about 239 to 281 ms one way from distance alone, roughly ten times the 20 to 30 msec at which a talker finds their own delayed voice disturbing (section 2). Broadcast-quality audio coding can add 170 to 400 ms on top. The same 2004 report notes that the two halves of a contribution loop are rarely identical: broadcasters commonly put a high-quality circuit on the programme line, a lower-quality one on the foldback line, and often send audio-only foldback even on video links.
The newer options cut the delay but add variability. Low Earth orbit (LEO) satellite service such as Starlink brought median latency down to around 80 ms, but with two distinct clusters of latency and packet loss that varied more than tenfold between neighbouring countries. Private 5G networks can carry production video with about 80 to 105 ms of total delay, while news crews using bonded cellular (sending one camera feed over several public mobile networks at once) accept up to two seconds and give up live two-way conversation as a result.
| Metric | Value | Source |
|---|---|---|
| Geostationary satellite path delay from distance alone: satellite directly overhead vs at the horizon | About 239 ms vs about 281 ms | ITU-R Report BT.2044 (2004) |
| Audio coding delay: broadcast-quality MPEG AAC vs low-delay AAC-LD | 170 to 400 ms vs around 60 ms | ITU-R Report BT.2044 (2004) |
| Starlink median TLS handshake latency in 2024 (RIPE Atlas, 146 probes), and the second latency cluster | Approximately 80 ms; second cluster at about 150 to 250 ms | Richter, Ververis and Bajpai, IFIP Networking 2025 |
| Starlink packet loss, January 2022 to June 2024: Germany vs Austria | 10.52% vs 0.73% | Richter, Ververis and Bajpai, IFIP Networking 2025 |
| Private 5G uplink transit time across 10 tested configurations (20 Mb/s unless stated): range of means vs range of maximums | 6.4 to 13.8 ms vs 10.2 to 53.6 ms | BBC R&D, University of Strathclyde and TV2, IBC2024 |
| Video latency, encoder input to decoder output: private 5G typical vs optimized, compared with DVB-T COFDM wireless camera links | Around 100 to 105 ms vs 80 ms vs typically around 50 ms | BBC R&D, University of Strathclyde and TV2, IBC2024 |
| Video delay tolerated by news contribution bonded over public mobile networks | Up to 2 seconds | BBC R&D, University of Strathclyde and TV2, IBC2024 |
Outlier note: the Starlink figures are TLS handshake times to root servers, not a one-way audio path, and the authors note that their result contradicts earlier studies reporting medians well below 80 ms. The 5G results come from single-cell private networks, and the authors add that a cell handover interrupts data for around 20 ms. Most recent available data for satellite and codec delay: Report BT.2044, 2004.
6. Remote Production and Cloud Adoption
The adoption numbers explain why the latency numbers matter more each year. Remote production (often called REMI, for remote integration model: cameras and microphones at the venue, the gallery and crew somewhere else) moves the talkback loop from a cable run inside a truck to a wide-area network. 41% of more than 1,300 surveyed professionals named remote production as their biggest technology priority for 2026, and the share using cellular networks for live contribution rose from 46% to 54% in one year.
The largest live productions have already crossed over. At Paris 2024, two-thirds of booked remote services ran over the cloud, taking over from satellite, which had been introduced for the Tokyo 1964 Games. The survey data needs care, though: both surveys below draw on self-selected respondents, and the larger one is run by a vendor of contribution equipment.
| Metric | Value | Source |
|---|---|---|
| Respondents naming remote production as their biggest technology priority for 2026 (more than 1,300 respondents) | 41% | Haivision Broadcast Transformation Report 2026, as reported by SVG, March 2026 (vendor survey) |
| Respondents using cellular networks for live video contribution, 2026 report vs 2025 report | 54% vs 46% | Haivision Broadcast Transformation Report 2026 (vendor survey) |
| Respondents using cellular networks as backup to fiber | 61% | Haivision Broadcast Transformation Report 2026 (vendor survey) |
| Respondents using the SRT transport protocol, 2020 report vs 2026 report | 47% vs 78% | Haivision Broadcast Transformation Report 2026 (vendor survey) |
| Broadcast and media professionals implementing cloud production tools (312 self-selected readers, November 2024 to January 2025) | 60% | NewscastStudio industry survey, February 2025 |
| Top barriers to cloud production: cost vs workflow complexity vs security concerns | 33% vs 25% vs 17% | NewscastStudio industry survey, 2025 |
| Paris 2024: share of booked remote services using OBS Live Cloud | Two-thirds, across 54 broadcasters | Olympic Broadcasting Services and Alibaba Cloud, as reported by TV Technology, 2024 |
| Feeds carried over OBS Live Cloud at Paris 2024, vs broadcasters subscribed when the service became standard at Beijing 2022 | 379 video (11 UHD, 368 HD) and 100 audio feeds vs 22 broadcasters | Olympic Broadcasting Services, as reported by TV Technology, 2024 |
Context note: the OBS share counts booked remote services, not all rights-holding broadcasters, and the figures were announced as the Games began rather than measured afterwards.
Summary: Broadcast Talkback and IFB Latency by the Numbers
| Metric | Value | Source |
|---|---|---|
| Detectability threshold for sound/picture timing error | About +45 ms to -125 ms | ITU-R BT.1359-1 (1998) |
| Acceptability threshold for sound/picture timing error | About +90 ms to -185 ms | ITU-R BT.1359-1 (1998) |
| End-to-end lip-sync limit at any output intended for emission (sound before vs after picture) | 40 ms or less vs 60 ms or less | EBU R37-2007 |
| Sound timing at DTV encoder inputs | Never more than 15 ms early or 45 ms late | ATSC IS-191 (reported 2004) |
| Own-voice delay rated disturbing while reading aloud (32 participants) | 25 to 30 msec | Stone and Moore, Ear and Hearing 2002 |
| Own-voice delay range rated probably just acceptable (25 hearing-impaired talkers) | 14 to 30 msec | Stone and Moore, Ear and Hearing 2005 |
| Threshold for noticing a delayed own voice (6 participants) | Approximately 100 ms (SD 66.54 ms) | Yamamoto and Kawabata, PLoS ONE 2011 |
| Processing delay tested on 20 normally hearing and 20 hearing-impaired participants | Up to 50 ms | Goehring et al., International Journal of Audiology 2018 |
| In-ear monitor latency target for performers | 4 ms | BBC R&D, University of Strathclyde and TV2, IBC2024 |
| DECT talkback latency: to beltpack vs beltpack to beltpack | Below 40 ms vs below 70 ms | BBC R&D, University of Strathclyde and TV2, IBC2024 |
| Wireless beltpack audio latency, manufacturer’s own claim | Under 35 ms | Clear-Com FSE-BP50 datasheet, 2023 |
| ST 2110-30 packet time: Level A vs Levels B and C | 1 ms vs 125 microseconds | AIMS, AES67 and ST 2110-30 introduction |
| Starlink median TLS handshake latency, 2024 | Approximately 80 ms | Richter, Ververis and Bajpai, IFIP Networking 2025 |
| Starlink packet loss in Germany, January 2022 to June 2024 | 10.52% | Richter, Ververis and Bajpai, IFIP Networking 2025 |
| Remote production named as biggest technology priority for 2026 | 41% | Haivision Broadcast Transformation Report 2026 (vendor survey) |
| Professionals implementing cloud production tools | 60% | NewscastStudio industry survey, 2025 |
| Booked remote services on OBS Live Cloud at Paris 2024 | Two-thirds, across 54 broadcasters | Olympic Broadcasting Services, 2024 |
Methodology and Sources
- International Telecommunication Union: Recommendation BT.1359-1, Relative timing of sound and vision for broadcasting (1998); Report BT.2044, Tolerable round-trip time delay for sound-programme and television broadcast programme inserts (2004); Recommendation G.114, One-way transmission time (2003).
- European Broadcasting Union: R37-2007, The relative timing of the sound and vision components of a television signal.
- ATSC Implementation Subcommittee finding IS-191, as reported by TV Technology, ATSC Recommends AV Sync Rules, January 7, 2004.
- Stone and Moore, Tolerable hearing aid delays, Ear and Hearing: part I, 1999; part II, 2002; part IV, 2005.
- Yamamoto and Kawabata, Temporal Recalibration in Vocalization Induced by Adaptation of Delayed Auditory Feedback, PLoS ONE 2011.
- Goehring, Chapman, Bleeck and Monaghan, Tolerable delay for speech production and perception, International Journal of Audiology 2018.
- Waddell et al. (BBC R&D, University of Strathclyde and Neutral Wireless, TV2), Optimising 5G for low latency broadcast production, IBC2024.
- Clear-Com, FSE-BP50 FreeSpeak Edge beltpack datasheet, revision August 2023 (manufacturer’s own specification).
- AIMS (Alliance for IP Media Solutions), Introduction to AES67 and its relationship with SMPTE ST 2110-30, presented at the AES New York convention.
- Richter, Ververis and Bajpai (HPI, University of Potsdam), Breaking Through the Clouds: Performance Insights into Starlink’s Latency and Packet Loss, IFIP Networking 2025.
- Haivision, Broadcast Transformation Report 2026, as reported by Sports Video Group, March 4, 2026.
- NewscastStudio, industry survey, February 7, 2025, 312 respondents, conducted November 2024 to January 2025.
- Olympic Broadcasting Services and Alibaba Cloud, OBS Cloud 3.0 announcement, as reported by TV Technology, 2024.
- Data watch: most of the perceptual and standards data is old (1998 to 2018) because the underlying limits have not been retested; these are flagged as most recent available data in each section. ATSC’s implementation subcommittee judged the 1998 international lip-sync recommendation inadequate for DTV, so the two sets of figures are intentionally different, not a measurement conflict. The Haivision and NewscastStudio results come from self-selected samples, and the vendor describes the SRT rise from 47% to 78% as a ‘31% increase’, which is a difference in percentage points between two different samples. The Starlink medians measure TLS handshakes and conflict with earlier studies that reported lower medians. The beltpack figure from Clear-Com is a design-center specification, not an independent measurement. No public, independent dataset of measured IFB round-trip latency in live remote productions was found; we will add one if a broadcaster or standards body publishes it.
Last updated: November 8, 2026. We update this roundup quarterly, and the next refresh is expected when Haivision publishes its 2027 Broadcast Transformation Report.