The most quoted echo canceller number, ERLE, agrees with human listeners at a Pearson correlation of just 0.31, while Microsoft’s neural AECMOS metric reaches 0.99 on far-end echo ratings (Microsoft, ICASSP 2021 and ICASSP 2023 AEC Challenge papers). On the standards side, ITU-T G.168 gives a network echo canceller roughly 1 second to converge, 3GPP TS 26.131 asks headsets for at least 46 dB of echo loss, and Microsoft’s challenge halved the latency allowance from 40 ms to 20 ms between 2021 and 2023. Echo is also the problem that pushes people toward better room hardware, which we covered from the microphone side in our conference room ceiling beamforming statistics. We aggregated data from ITU-T, 3GPP, the four Microsoft Research AEC Challenge papers, Amazon Web Services, ByteDance, Alibaba’s Tongyi Lab, independent arXiv researchers, and HEAD acoustics.
TL;DR
- Network echo cancellers must reach their required combined loss within about 1 second under Test 2A (ITU-T G.168, 04/2015).
- Headsets need at least 46 dB of weighted terminal coupling loss at any volume, with 55 dB recommended at nominal volume (3GPP TS 26.131, Rel-19).
- ERLE correlates with listener ratings at PCC 0.31 and SRCC 0.23 (Microsoft, ICASSP 2021 AEC Challenge paper).
- AECMOS correlates with far-end echo ratings at PCC 0.99 (Microsoft, ICASSP 2023 AEC Challenge paper).
- The challenge latency cap fell from 40 ms algorithmic latency in 2021 to 20 ms algorithmic plus buffering in 2023 (Microsoft AEC Challenge papers and repository).
- The top-scoring INTERSPEECH 2021 entry averaged 4.43 against a 3.87 baseline (Microsoft, INTERSPEECH 2021 AEC Challenge results).
- Mobile devices scored 0.75 MOS lower than desktops on double-talk degradation (Microsoft, ICASSP 2022 AEC Challenge paper).
- WebRTC AEC3 removed about 30 dB of echo versus 40.8 dB for a neural suppressor (ByteDance, arXiv 2020).
- A January 2026 end-to-end model reached 78.69 dB ERLE with 1.2 million parameters (Alibaba Tongyi Lab, E2E-AEC arXiv 2026).
- A small neural echo and noise model runs in 0.14 ms per frame, a real-time factor of 0.014 (Microsoft, DeepVQE Interspeech 2023).
- Hands-free terminals in the best double-talk category (category 1) attenuate sent speech by at most 3 dB during double talk (HEAD acoustics, describing ITU-T P.340).
1. The ITU-T Rulebook: Convergence, Double Talk and Leak Tests
The core telecom rules for echo cancellers are old but still in force. “Echo” here is your own voice coming back to you after it has travelled to the far end and leaked into the return path. G.168 gives a canceller roughly one second to reach its required loss, then keeps testing whether it holds that loss when both people talk at once (“double talk”) and when the line goes quiet for minutes. A non-linear processor (NLP) is the final stage that suppresses whatever residual echo the adaptive filter leaves behind; several tests switch it off so the filter itself is measured.
The real-world echo paths these tests model are short. In field data reproduced in ITU-T G.168 (04/2015), 101 long-distance calls from Montreal across North America (June 1998 to April 1999) were measured, the histogram of 88 of them showed a mean dispersion time of 6.02 ms, and no call exceeded 12 ms. The acoustic counterpart, G.167, has been withdrawn and its content moved into P.340 and G.161.
| Metric | Value | Source |
|---|---|---|
| Edition of the network echo canceller Recommendation in force | G.168 (04/2015), plus Corrigendum 1 (12/2022) | ITU-T G.168 |
| Minimum echo return loss (the loss the echo path itself provides) used in tests; number of standard echo path models | 6 dB minimum; 8 models (m1 to m8), with minimums of 6.55 dB, 11.06 dB and 9.27 dB for m2, m7 and m8 | ITU-T G.168 (04/2015), Annex D |
| Test 2A convergence, non-linear processor enabled | Required combined loss within 1 s plus echo path delay and start offset; peaks no more than 5 dB above the limit | ITU-T G.168 (04/2015) |
| Test 3A, double talk with a near-end signal 15 dB below the far-end level (far-end -25 to 0 dBm0) | Must converge within 5 s | ITU-T G.168 (04/2015) |
| Test 3B(a) and Test 4: residual echo rise allowed during double talk, and 2 minutes after the far-end signal stops (leak test) | No more than 10 dB above the steady-state requirement in both cases | ITU-T G.168 (04/2015) |
| Status of G.167, the original acoustic echo controller Recommendation (03/1993) | Withdrawn; replaced by P.340 (2000) for clauses 1-5 and G.161 (2002) for clause 6 | ITU-T G.167 record |
| Latest edition of the hands-free terminal Recommendation covering echo and double-talk performance | P.340 (07/2026) | ITU-T P.340 summary 2026 |
| Frequency range of the in-vehicle wideband hands-free Recommendation | 50 to 7,000 Hz; edition in force P.1110 (10/2025) | ITU-T P.1110 |
Context note: dBm0 is a signal level relative to a reference point in the network, so “-25 to 0 dBm0” means a quiet to loud far-end talker. The P.340 2026 revision is new; its full text was not public at the time of writing, so this roundup quotes only its published summary.
2. Echo Loss Targets for Headsets, Hands-Free Devices and Networks
Terminal coupling loss (TCLw) is how much of the far-end voice, played through a device’s speaker or earpiece, is attenuated before it leaks back into that device’s own microphone and is sent back. The mobile standards set 46 dB as the floor for headsets at any volume, and the recommended target of 55 dB comes with an admission that it may not be measurable on every device because of the device’s own idle noise.
The interesting gap is hands-free. A desktop speakerphone may drop to 40 dB at maximum volume and still comply, which is why hands-free double-talk behaviour gets its own categories: a category 1 device barely ducks your voice while the other side talks (at most 3 dB), while a category 3 device cuts it by more than 12 dB, close to walkie-talkie behaviour where only one side is heard at a time.
| Metric | Value | Source |
|---|---|---|
| Headset weighted terminal coupling loss: at any volume setting vs recommended at nominal volume | At least 46 dB vs at least 55 dB | 3GPP TS 26.131 (Rel-19) |
| Desktop hands-free weighted terminal coupling loss: any volume vs nominal volume in free-field conditions | At least 40 dB vs at least 46 dB | 3GPP TS 26.131 (Rel-19) |
| Desktop hands-free recommended performance objective at nominal volume | At least 55 dB (may not be measurable because of idle channel noise) | 3GPP TS 26.131 (Rel-19) |
| Echo loss presented by GSM, 3G, LTE or NR networks at the point of interconnect, single talk | At least 46 dB | 3GPP TS 26.132 (Rel-19) |
| Speech discarded at the start of the echo loss test so the echo canceller can converge; test signal level | First 17.0 s; 10 dBm0 | 3GPP TS 26.132 (Rel-19) |
| Terminal coupling loss guidance: digital input to output (per P.341) vs field conditions (per P.342) vs desirable for G.131 talker echo objectives | At least 35 dB vs greater than 40 dB vs greater than 45 dB | ITU-T G.168 (04/2015), Appendix I |
| Double-talk categories for hands-free terminals and attenuation of sent speech | 5 categories (1, 2a, 2b, 2c, 3): at most 3 dB for category 1, at most 9 dB for category 2b, more than 12 dB for category 3 | HEAD acoustics ACOPT 32 datasheet, describing ITU-T P.340 |
Context note: the 3GPP values were read from the Rel-19 text of TS 26.131 and TS 26.132; TS 26.132 has a newer Rel-20 file dated July 2026 in the 3GPP archive. The HEAD acoustics document is a test-equipment datasheet that restates the P.340 categories; it does not give the 2a and 2c limits.
3. Latency Budgets: How Many Milliseconds an Echo Canceller May Spend
Every millisecond an echo canceller spends buffering audio is a millisecond added to the conversation. The research benchmark moved hard on this: Microsoft’s challenge allowed 40 ms of algorithmic latency in 2021 and halved the allowance to 20 ms of algorithmic plus buffering latency in 2023, and the organizers’ own model lost essentially nothing in echo scores at the tighter limit. The ICASSP 2023 paper names an end-to-end target below 50 ms (the just-noticeable difference) and a future goal of under 5 ms for algorithm plus buffering (Microsoft, ICASSP 2023 AEC Challenge paper).
Compute is the other budget. A neural model that needs a fraction of a millisecond per frame can run on every laptop in a meeting; one that needs most of a CPU core cannot. For how the rest of the delay budget is spent by jitter buffers and codecs, see our VoIP jitter buffer and speech quality statistics.
| Metric | Value | Source |
|---|---|---|
| One-way mouth-to-ear delay for essentially transparent interactivity vs ceiling for general network planning | Below 150 ms vs 400 ms | ITU-T G.114 (05/2003) |
| AEC Challenge algorithmic latency limit, ICASSP 2021 | At most 40 ms (a 20 ms frame with a 10 ms stride, 30 ms, was allowed) | Microsoft, ICASSP 2021 AEC Challenge paper |
| AEC Challenge latency limit, ICASSP 2023 | 20 ms of algorithmic plus buffering latency | Microsoft AEC-Challenge repository |
| AEC Challenge compute limit, ICASSP 2023 | Real-time factor of at most 0.5 on an Intel Core i5 quad-core at 2.4 GHz, single thread | Microsoft, ICASSP 2023 AEC Challenge paper |
| Microsoft model at 20 ms vs 2022 challenge winner at 40 ms: far-end echo MOS; double-talk echo MOS | 4.65 vs 4.66; 4.68 vs 4.68 | Microsoft, ICASSP 2023 AEC Challenge paper |
| DeepVQE full model vs DeepVQE-S small model: parameters and CPU time per frame | 7.5M and 3.66 ms vs 0.59M and 0.14 ms (real-time factor 0.014) | Microsoft, DeepVQE, Interspeech 2023 |
| Combined linear echo canceller plus neural residual echo suppressor, 16 kHz challenge entry | 5.5% of one Intel i7-8565U core, 0.55 ms per 10 ms frame | Amazon Web Services, PercepNet echo control, arXiv 2021 |
Most recent available data for the challenge limits: Microsoft, 2023. The AEC-Challenge repository still lists ICASSP 2023 as the latest edition, so these remain the latest public benchmark rules from that series.
4. AEC Challenge Scoreboards, 2021 to 2023
Microsoft’s AEC Challenge is the closest thing the field has to a public league table. Entries are rated by crowdsourced listeners on a 1 to 5 scale (MOS for near-end speech quality, DMOS, a degradation score, for echo and other artefacts). At INTERSPEECH 2021 the top-scoring entry averaged 4.43 against a 3.87 baseline (the top four were statistically tied, so Microsoft ranked them by computational complexity), and almost all of the gain came from echo removal and double talk (the scenario where older cancellers either let echo through or chop the near-end talker), not from near-end speech quality, which moved only from 4.18 to 4.25.
The datasets grew with each round: more than 2,500 real environments and devices for ICASSP 2021 (Microsoft, ICASSP 2021 AEC Challenge paper), 5,000 for INTERSPEECH 2021, and more than 10,000 for the 2023 round (the 2022 count differs between Microsoft’s papers, see Methodology). Public data on how often echo actually reaches users in production calls is still missing: Microsoft’s papers describe echo only as “one of the major sources of poor speech quality ratings”, without a percentage.
| Metric | Value | Source |
|---|---|---|
| ICASSP 2021: baseline vs first-place system, mean of the four subjective tests | 3.68 vs 4.11 (17 submissions) | Amazon Web Services, PercepNet echo control, arXiv 2021 |
| ICASSP 2021: baseline vs first-place system, degradation of near-end speech during double talk (DMOS) | 3.28 vs 4.07 | Amazon Web Services, PercepNet echo control, arXiv 2021 |
| INTERSPEECH 2021: top-scoring entry (ERCESI) vs baseline, mean score | 4.43 vs 3.87 | Microsoft, INTERSPEECH 2021 AEC Challenge results |
| INTERSPEECH 2021: top-scoring entry vs baseline, far-end single-talk echo DMOS; double-talk other degradations DMOS | 4.59 vs 3.82; 4.18 vs 3.45 | Microsoft, INTERSPEECH 2021 AEC Challenge results |
| INTERSPEECH 2021: ranked team entries with a mean above the baseline | 10 of 14 (counted from the results table) | Microsoft, INTERSPEECH 2021 AEC Challenge results |
| ICASSP 2022: submissions; mobile vs desktop gap on double-talk degradation | 18 submissions; mobile lower by 0.75 MOS | Microsoft, ICASSP 2022 AEC Challenge paper |
| ICASSP 2023: entries; winning score and model size | 20 entries (17 non-personalized, 3 personalized); 0.854 with 1.5M parameters | Microsoft, ICASSP 2023 AEC Challenge paper |
| ICASSP 2023: baseline vs winner, final score and speech recognition word accuracy | 0.736 vs 0.852; 64.9% vs 82.3% | Microsoft, DeepVQE, Interspeech 2023 |
Context note: the 2023 final score is a 0 to 1 composite that averages five rescaled subjective scores with word accuracy, so it is not comparable with the 2021 MOS means. The two Microsoft papers print slightly different winning scores for 2023 (see Methodology). Most recent available data: Microsoft, 2023.
5. Measuring Echo: Why ERLE Lost Ground to Perceptual Metrics
ERLE (echo return loss enhancement) is the number of decibels an echo canceller removes from the microphone signal. It is easy to compute, which is why it still headlines datasheets, but ERLE correlates with listener ratings at only 0.31 (Pearson), because a canceller can score a huge ERLE by muting the line and destroying the near-end talker’s voice. PESQ, an older intrusive speech-quality metric that compares the output with a clean reference, does better at 0.67.
AECMOS, Microsoft’s neural estimator, needs no clean reference and predicts the echo and degradation scores listeners would give. The practical change: teams can now rank models without paying for a listening test every time. That matters for anyone who has ever compared devices in a lab instead of in real rooms, a gap we covered in our anechoic chamber acoustic testing statistics.
| Metric | Value | Source |
|---|---|---|
| ERLE vs subjective ratings: Pearson (PCC) and Spearman (SRCC) correlation | 0.31 and 0.23 | Microsoft, ICASSP 2021 AEC Challenge paper |
| PESQ vs subjective ratings: PCC and SRCC | 0.67 and 0.57 | Microsoft, ICASSP 2021 AEC Challenge paper |
| AECMOS (2023 version), Pearson correlation with listener ratings: far-end single-talk echo; double-talk echo; double-talk other degradations | 0.99 (SRCC 0.92); 0.99; 0.96 | Microsoft, ICASSP 2023 AEC Challenge paper |
| Original AECMOS training set; crowdsourced votes per clip | 64,013 clips (145.8 hours); 5 votes per clip | Microsoft, AECMOS paper, ICASSP 2022 |
| Original AECMOS stack ranking: models ranked; per-contestant PCC on far-end single-talk echo | 22 models (14 contestants, 4 deep and 4 DSP models from the authors); 0.996 | Microsoft, AECMOS paper, ICASSP 2022 |
| Speech recognition word accuracy vs mean subjective score across ICASSP 2022 entries (PCC) | 0.85 | Microsoft, ICASSP 2022 AEC Challenge paper |
| Listeners per utterance in the ICASSP 2021 subjective tests; 95% confidence interval | 10 listeners; 0.01 MOS | Amazon Web Services, PercepNet echo control, arXiv 2021 |
Outlier note: per-clip correlations are much weaker than per-model ones. The original AECMOS paper reports a per-clip PCC of 0.582 for double-talk echo, so the metric is reliable for ranking systems, not for judging a single recording.
6. Neural vs Linear Echo Control: What the Head-to-Heads Show
Linear adaptive filters model the echo path as a straight-line system, so they struggle with echo distorted by a cheap, overdriven laptop speaker. In the ICASSP 2022 top five, most neural systems were built on top of a linear filter rather than replacing it. In one ByteDance test, WebRTC AEC3 removed about 30 dB of echo while a neural residual suppressor reached 40.8 dB, and AEC3 actually made double-talk speech quality worse than doing nothing at a 0 dB signal-to-echo ratio (echo as loud as the near-end voice).
Treat the ERLE figures with the caution explained in section 5: a 78.69 dB ERLE says the line was very quiet during far-end speech, not that the near-end voice survived. On the vendor side, Microsoft says Teams’ echo model was trained on 30,000 hours of speech samples (Microsoft 365 Blog 2022, the vendor’s own claim), and its DeepVQE researchers say the small model was tested in Teams for hundreds of millions of users.
| Metric | Value | Source |
|---|---|---|
| ERLE, far-end single talk with speech: linear adaptive filter vs WebRTC AEC3 vs adaptive filter plus neural residual suppressor | 16.612 dB vs 29.976 dB vs 40.786 dB | ByteDance, residual echo suppression, arXiv 2020 |
| ERLE, far-end single talk with speech plus music: WebRTC AEC3 vs neural residual suppressor | 25.376 dB vs 43.144 dB | ByteDance, residual echo suppression, arXiv 2020 |
| PESQ during double talk at 0 dB signal-to-echo ratio: unprocessed vs WebRTC AEC3 vs neural residual suppressor | 1.908 vs 1.610 vs 2.816 | ByteDance, residual echo suppression, arXiv 2020 |
| ERLE: Speex vs WebRTC vs adaptive filter combined with a recurrent neural network | 25 dB vs 40 dB vs 68 dB | Ma et al., arXiv 2020 |
| Dual-signal LSTM echo canceller (DTLN-aec): MOS gain over the AEC Challenge baseline; training data | 0.30 MOS; 60 hours of real and synthetic echo scenarios | Westhausen and Meyer, DTLN-aec, arXiv 2020 |
| ICASSP 2022 top 5 systems: hybrid linear plus neural designs; parameter range | 3 of 5; 1.5M to 55.5M parameters | Microsoft, ICASSP 2022 AEC Challenge paper |
| ICASSP 2023 entries: correlation of final score with model size; with using a hybrid design (PCC) | -0.54; 0.10 | Microsoft, ICASSP 2023 AEC Challenge paper |
| End-to-end neural echo canceller (January 2026): ERLE with voice activity masking; parameters; MOS average gain from one training strategy | 78.69 dB; 1.2M parameters; 4.31 to 4.38 | Alibaba Tongyi Lab, E2E-AEC, arXiv 2026 |
Context note: these results come from different test sets, so they rank methods within each paper, not across papers. Most recent available data for the AEC3 and Speex comparisons: ByteDance and Ma et al., 2020.
Summary: Acoustic Echo Cancellation by the Numbers
| Metric | Value | Source |
|---|---|---|
| G.168 Test 2A convergence time | About 1 s plus echo path delay | ITU-T G.168 (04/2015) |
| Headset weighted terminal coupling loss, any volume vs nominal | At least 46 dB vs at least 55 dB | 3GPP TS 26.131 (Rel-19) |
| Network echo loss at the point of interconnect | At least 46 dB | 3GPP TS 26.132 (Rel-19) |
| Hands-free double-talk categories | 5; category 1 attenuates sent speech by at most 3 dB | HEAD acoustics, describing ITU-T P.340 |
| Hands-free terminal Recommendation, latest edition | P.340 (07/2026) | ITU-T P.340 summary 2026 |
| One-way delay for transparent interactivity | Below 150 ms | ITU-T G.114 (05/2003) |
| AEC Challenge latency cap, 2023 | 20 ms algorithmic plus buffering | Microsoft AEC-Challenge repository |
| ERLE vs listener ratings | PCC 0.31 | Microsoft, ICASSP 2021 AEC Challenge paper |
| AECMOS vs far-end echo ratings | PCC 0.99 | Microsoft, ICASSP 2023 AEC Challenge paper |
| Original AECMOS training set | 64,013 clips, 145.8 hours | Microsoft, AECMOS paper, ICASSP 2022 |
| ICASSP 2021 baseline vs winner | 3.68 vs 4.11 | Amazon Web Services, PercepNet echo control, arXiv 2021 |
| INTERSPEECH 2021 top-scoring entry vs baseline | 4.43 vs 3.87 | Microsoft, INTERSPEECH 2021 AEC Challenge results |
| Mobile vs desktop, double-talk degradation | Mobile lower by 0.75 MOS | Microsoft, ICASSP 2022 AEC Challenge paper |
| DeepVQE-S CPU time per frame | 0.14 ms (real-time factor 0.014) | Microsoft, DeepVQE, Interspeech 2023 |
| WebRTC AEC3 vs neural residual suppressor, ERLE | 29.976 dB vs 40.786 dB | ByteDance, residual echo suppression, arXiv 2020 |
| Speex vs WebRTC vs hybrid RNN, ERLE | 25 dB vs 40 dB vs 68 dB | Ma et al., arXiv 2020 |
| DTLN-aec gain over the challenge baseline | 0.30 MOS | Westhausen and Meyer, DTLN-aec, arXiv 2020 |
| End-to-end neural AEC, ERLE (2026) | 78.69 dB with 1.2M parameters | Alibaba Tongyi Lab, E2E-AEC, arXiv 2026 |
Methodology and Sources
- ITU-T: G.168, Digital network echo cancellers (04/2015, Corrigendum 12/2022), G.167, Acoustic echo controllers (03/1993, withdrawn), P.340 summary (07/2026), P.1110, Wideband hands-free communication in motor vehicles (10/2025), G.114, One-way transmission time (05/2003)
- 3GPP: TS 26.131, Terminal acoustic characteristics for telephony; Requirements (Rel-19) and TS 26.132, Speech and video telephony terminal acoustic test specification (Rel-19)
- HEAD acoustics: ACOPT 32 Speech-based Double Talk datasheet (restates the ITU-T P.340 categories)
- Microsoft Research: ICASSP 2021 AEC Challenge paper, INTERSPEECH 2021 AEC Challenge results, ICASSP 2022 AEC Challenge paper, ICASSP 2023 AEC Challenge paper, AECMOS paper, DeepVQE, Interspeech 2023, AEC-Challenge repository
- Amazon Web Services: Low-complexity joint neural echo control and speech enhancement based on PercepNet (2021)
- ByteDance: Residual acoustic echo suppression based on efficient multi-task convolutional neural network (2020)
- Ma, Huang, Zhao and Su: Acoustic echo cancellation by combining adaptive digital filter and recurrent neural network (2020)
- Westhausen and Meyer: Acoustic echo cancellation with the dual-signal transformation LSTM network (2020)
- Alibaba Tongyi Lab: E2E-AEC, an end-to-end neural network approach for acoustic echo cancellation (January 2026)
- Microsoft 365 Blog: How Microsoft Teams uses AI and machine learning to improve calls and meetings (2022, vendor claim)
- Data watch: no platform has published a telemetry-based rate of echo in real calls, so this roundup reports none; Microsoft’s papers describe echo only qualitatively as a top complaint. The two Microsoft sources disagree slightly on the 2023 winner: the ICASSP 2023 challenge paper lists the first-place score as 0.854, while the DeepVQE paper’s table lists the AEC winner at 0.852. Microsoft’s sources also disagree on the ICASSP 2022 dataset: the ICASSP 2022 paper says more than 10,000 environments and devices, while the ICASSP 2023 paper lists 7,500 for that round. The ERLE 0.31 and PESQ 0.67 correlations are the same in the 2021, 2022 and 2023 challenge papers, which suggests one underlying study rather than three independent measurements. AECMOS correlations vary by model version (the ICASSP 2022 paper reports 0.828 PCC for far-end echo, the 2023 paper 0.99). The INTERSPEECH 2021 paper says nearly all participants beat the baseline, while the public results table shows 10 of 14 ranked entries above the baseline mean. The ERLE values in the 2020 Speex and WebRTC comparisons come from each paper’s own test set and settings. Most AEC Challenge data is from 2021 to 2023 because the AEC-Challenge repository lists ICASSP 2023 as the latest edition; the P.340 (07/2026) full text and the 3GPP TS 26.132 Rel-20 file are the next updates to check.
Last updated: October 24, 2026. We update this roundup quarterly, and the next refresh is expected when the full text of ITU-T P.340 (07/2026) is published.