The best neural packet loss concealment system in Microsoft’s INTERSPEECH 2022 Audio Deep PLC Challenge scored a PLCMOS of 4.28 on a 1 to 5 scale, against 2.90 for simply inserting silence and 4.51 for the clean original (Microsoft, INTERSPEECH 2022 Audio Deep PLC Challenge paper). Two years later, when the same challenge moved to full-band audio and loss bursts of up to 3 seconds, the shared winners could only lift overall quality from 2.37 to 3.49, still well below the 4.01 that clean speech earned (Microsoft, ICASSP 2024 Audio Deep PLC Challenge paper). The problem is not rare: Google reported that 99% of Duo calls had to deal with packet loss, jitter or delay, and one call in ten lost more than 8% of its audio (Google Research, WaveNetEQ blog 2020). Concealment is the quiet layer underneath the business trends in our VoIP and UCaaS statistics, and it complements the operator-side picture in our cellular voice quality (MOS) statistics. We aggregated data from Microsoft Research challenge papers, the IETF and ITU-T, the W3C, Google Research, Xiph.Org’s Opus developers, Meta Engineering, Microsoft Learn and Twilio documentation, and peer-reviewed PLC studies listed in the methodology.
TL;DR
- 99% of Google Duo calls had to handle packet loss, jitter or delay; 10% lost more than 8% of total audio (Google Research, WaveNetEQ blog 2020).
- Microsoft Teams marks an audio stream Poor when average packet loss exceeds 0.1, meaning 10% (Microsoft Learn, CQD Stream Classification 2026).
- G.711 with its Appendix I concealment has a packet-loss robustness factor of 25.1, versus 4.3 with no concealment (ITU-T G.113, 09/2024).
- Neural PLC winner: PLCMOS 4.28 vs 2.90 for zero-fill and 4.51 for clean speech (Microsoft, INTERSPEECH 2022 Deep PLC Challenge paper).
- The Opus team’s hybrid PLC ranked first on word accuracy (0.882) and second overall in the 2022 challenge (Valin et al., Real-Time PLC with Mixed Generative and Predictive Model 2022).
- With bursts up to 3 seconds, 2 of 8 eligible systems scored below unprocessed lossy audio on P.804 overall quality (Microsoft, ICASSP 2024 Deep PLC Challenge paper).
- PLCMOS reaches a model-wise Pearson correlation of about 0.97 with human ratings; on reference-aligned data PESQ manages 0.52 (Diener et al., PLCMOS paper 2023).
- Opus LBRR in-band FEC covers 1 frame at about 2/3 of the regular bitrate; DRED covers up to 1 second at about 1/50 (IETF draft-ietf-mlcodec-opus-dred-07, August 2026).
- Opus needs at least 19 kbps to encode in-band FEC at 10% packet loss, according to Meta’s own MLow comparison (Meta Engineering, MLow 2024).
- Opus 1.6 shrank its DRED models about 3x, to 600 kB from 1,800 kB (Xiph.Org, Opus 1.6 demo, December 2025).
1. How Much Audio Gets Lost, and Where Platforms Draw the Line
A packet, in this context, is one small slice of audio (often 20 ms) sent over the network. When it arrives too late or not at all, the receiver has two choices: play silence, or conceal the gap by synthesizing something that sounds like the missing speech. Google’s own data showed that 99% of Duo calls had to deal with packet loss, jitter or delay (Google Research, Improving Audio Quality in Duo with WaveNetEQ 2020). Concealment is therefore not an edge case: it runs in almost every call.
Platforms disagree on where loss turns from background noise into a problem. Microsoft Teams draws its Poor line at 10% average loss, Twilio at 5%, and Microsoft explicitly warns that a Poor label does not mean users actually perceived a quality issue, because its media stack corrects for loss (Microsoft Learn, Stream Classification in CQD 2026). That caveat is the whole argument for better concealment: the further PLC stretches, the higher a platform can set its alarm threshold.
| Metric | Value | Source |
|---|---|---|
| Share of Google Duo calls that must deal with packet loss, jitter or delay | 99% | Google Research, WaveNetEQ blog 2020 |
| Share of Duo calls that lose more than 3% of their total audio duration | 20% | Google Research, WaveNetEQ blog 2020 |
| Share of Duo calls that lose more than 8% of their total audio duration | 10% | Google Research, WaveNetEQ blog 2020 |
| Average packet loss rate above which Teams classifies an audio stream as Poor | Above 0.1 (10%), applied only to streams above 500 packets | Microsoft Learn, CQD Stream Classification 2026 |
| Packet loss level that triggers Twilio’s high_packet_loss call tag | Above 5% | Twilio Voice Insights, Details: Call Tags |
| Estimated MOS below which Twilio tags a browser or mobile SDK call low_mos | Below 3.5 | Twilio Voice Insights, Details: Call Tags |
| Average loss rate in the Deep PLC Challenge test set used to evaluate Opus DRED | 18.4%, with bursts up to one second | Valin et al., DRED paper (arXiv 2212.04453 v3, 2024) |
Most recent available data: Google Research, 2020. No major platform has published a newer call-level loss distribution. Teams reports the related “Ratio Concealed Samples Avg” per stream, where 0.1 means 10% of audio frames contained concealed samples (Microsoft Learn, CQD dimensions and measures 2026), but Microsoft does not publish fleet-wide values. The browser-side equivalents, concealedSamples and concealmentEvents, are defined in the W3C WebRTC Statistics API (Candidate Recommendation Draft, 25 September 2025).
2. Classical Codec Concealment Under the ITU-T E-model
Before neural networks, concealment meant repeating or extrapolating the last good pitch period. The ITU-T planning model (the E-model) scores how well a codec survives loss with the packet-loss robustness factor, Bpl: a plain number where higher means quality decays more slowly as packets disappear. G.711 with the concealment algorithm from its Appendix I has a Bpl of 25.1, almost six times the 4.3 of G.711 with no concealment (ITU-T G.113 09/2024, Table I.4).
What that gap means in practice is clearer with numbers. Using Equation 7-29 of ITU-T G.107 (06/2015), Ie-eff = Ie + (95 - Ie) x Ppl / (Ppl/BurstR + Bpl), G.711 (Ie = 0) at 5% random loss (BurstR = 1) gives 95 x 5 / (5 + 4.3) = 51.1 impairment points without concealment, and 95 x 5 / (5 + 25.1) = 15.8 with it. Subtracted from the default rating of R = 93.2, that is R 42.1 versus R 77.4, which the G.107 Annex B formula converts to an estimated MOS of about 2.2 versus 3.9 (our calculation; every other E-model input left at its default). Concealment alone moves a 5%-loss call from poor to acceptable.
| Metric | Value | Source |
|---|---|---|
| G.711 (64 kbit/s, 10 ms packets) with no concealment: packet-loss robustness factor Bpl under random loss | 4.3 | ITU-T G.113 (09/2024), Table I.4 |
| G.711 with the G.711 Appendix I concealment algorithm: Bpl under random loss | 25.1 | ITU-T G.113 (09/2024), Table I.4 |
| G.729 Annex A with voice activity detection (8 kbit/s, 20 ms packets, native concealment): Ie and Bpl | Ie 11, Bpl 19.0 | ITU-T G.113 (09/2024), Table I.4 |
| G.711 with 20 ms packets, 8 packets lost in one burst at 2% loss (burst ratio 7.84): effective impairment Ie-eff | 10 | ITU-T G.113 (09/2024), Table I.6 |
| G.722 at 64 kbit/s with the high-quality concealment of G.722 Appendix III: wideband Bpl (both ears listening) | 7.1 | ITU-T G.113 (09/2024), Appendix IV |
| EVS in super-wideband mode at 24.4 kbit/s: fullband Bpl | 11.4 | ITU-T G.113 (09/2024), Appendix V |
| E-model rating factor R when every input is at its default value | 93.2 | ITU-T G.107 (06/2015) |
Context: Bpl values are only comparable within the same scale (narrowband, wideband or fullband). G.113 lists no Opus values at all, so Opus-based WebRTC audio has to be judged with other tools (see section 5).
3. Neural Concealment vs Silence: The 2022 Deep PLC Challenge
The INTERSPEECH 2022 challenge benchmarked neural PLC on real loss patterns, using traces recorded from Microsoft Teams calls and a hard real-time rule: total algorithmic latency of at most 20 ms. Its headline result is the size of the gap between the winner and doing nothing. The winning system scored -0.55 on the comparison scale (CMOS, where 0 means indistinguishable from the clean original), versus -1.23 for zero-fill (Microsoft, INTERSPEECH 2022 Audio Deep PLC Challenge paper).
Intelligibility tells a more modest story. Word accuracy, measured as 1 minus the word error rate of a speech recognizer, rose only from 0.86 with silence to 0.88 for the winner, against 0.97 for clean speech. Listeners hear concealment as a large improvement, but the recognizer still missed 12% of words with the winning system, versus 3% on clean speech, a pattern worth keeping in mind next to our speech recognition word error rate statistics.
| Metric | Value | Source |
|---|---|---|
| Clean reference speech: PLCMOS (1 to 5) and word accuracy | 4.51 and 0.97 | Microsoft, INTERSPEECH 2022 Deep PLC Challenge paper |
| Zero-fill baseline (silence in place of lost packets): PLCMOS, CMOS and word accuracy | 2.90, -1.23 and 0.86 | Microsoft, INTERSPEECH 2022 Deep PLC Challenge paper |
| Winning system (team #12, time-domain U-Net, about 2.36M parameters): PLCMOS, CMOS and word accuracy | 4.28, -0.55 and 0.88 | Microsoft, INTERSPEECH 2022 Deep PLC Challenge paper |
| Lowest-ranked eligible system (team #6) vs zero-fill baseline: CMOS | -1.31 vs -1.23 | Microsoft, INTERSPEECH 2022 Deep PLC Challenge paper |
| Opus team’s hybrid generative and predictive PLC: word accuracy and rank on that metric | 0.882, first; second on CMOS and overall | Valin et al., Real-Time PLC with Mixed Generative and Predictive Model 2022 |
| Same system: computation per 10 ms of speech on a laptop CPU (Intel i7-10810U) | 1.3 to 1.4 ms, about 13% to 14% of one core | Valin et al. 2022 |
Outlier note: one ranked system scored below the silence baseline on CMOS, a reminder that a badly tuned generator can sound worse than a gap. The blind set was weighted toward short bursts: 504 of its 966 clips had a maximum burst of up to 120 ms. The hybrid system’s figures come from Valin et al., whose concealment later shipped in Opus 1.5 (see section 7).
4. Long Bursts: Where Concealment Still Breaks (ICASSP 2024)
The 2024 edition made the task deliberately harder: full-band 48 kHz audio, rated with the multi-dimensional ITU-T P.804 method (Microsoft Research, ICASSP 2024 Audio Deep PLC Challenge), and loss bursts of up to 3 seconds. Even the best system closed only about two thirds of the distance between lossy and clean audio: overall quality rose from 2.37 to 3.49, against 4.01 for clean speech ((3.49 - 2.37) / (4.01 - 2.37) = 68%) (Microsoft, ICASSP 2024 Audio Deep PLC Challenge paper).
The most telling column is discontinuity, P.804’s score for audible breaks and glitches (higher means smoother). It rose from 2.47 to 3.90 for the best system, a bigger jump than any other dimension. Concealment fixes the holes listeners notice first, while the synthesized speech still costs naturalness. Word accuracy shows the trade-off too: one co-winner recognized 0.81 of words, slightly below the 0.83 of the untouched lossy audio, a hint that invented speech can mislead a recognizer more than silence does.
| Metric | Value | Source |
|---|---|---|
| P.804 overall quality (1 to 5): clean speech vs unprocessed lossy audio vs the shared winners (1024K and NWPU & ByteAudio) | 4.01 vs 2.37 vs 3.49 and 3.44 | Microsoft, ICASSP 2024 Deep PLC Challenge paper |
| P.804 discontinuity score: clean vs lossy vs best system (1024K) | 4.55 vs 2.47 vs 3.90 | Microsoft, ICASSP 2024 Deep PLC Challenge paper |
| Word accuracy: clean vs lossy vs 1024K vs NWPU & ByteAudio | 0.98 vs 0.83 vs 0.81 vs 0.84 | Microsoft, ICASSP 2024 Deep PLC Challenge paper |
| Eligible systems scoring below unprocessed lossy audio on P.804 overall quality | 2 of 8 (2.11 and 2.17) | Microsoft, ICASSP 2024 Deep PLC Challenge paper |
| Audio bandwidth and listening test method of the 2024 edition | Full-band, 48,000 Hz; ITU-T P.804 multi-dimensional rating plus word error rate | Microsoft Research, ICASSP 2024 Deep PLC Challenge page |
Context: a ninth system scored 2.91 overall but missed the latency limit and was listed as not finished. The organizers note that 2022-era systems “may not be able to perform well for longer sequences of losses”, which is exactly where redundancy coding (section 6) takes over.
5. Scoring Concealment: Why Generic Quality Metrics Miss It
Concealment errors are short, intermittent and often plausible, which is precisely what general-purpose quality metrics were not built to catch. On reference-aligned data, PESQ reached a model-wise Pearson correlation of only 0.52 with human ratings of PLC systems, and STOI just 0.21, while PLCMOS reached 0.98 (Diener et al., PLCMOS paper, Microsoft, Interspeech 2023). A metric that ranks systems wrongly can steer a whole research field in the wrong direction.
PLCMOS works without a clean reference, which matters for real calls where the original audio is never available at the receiver. Its weakness is file-level precision: 0.85 per clip versus 0.97 per system, so it ranks algorithms far better than it judges a single call. The same tension runs through synthetic speech scoring, covered in our synthetic speech naturalness MOS statistics.
| Metric | Value | Source |
|---|---|---|
| PLCMOS vs human ratings: model-wise Pearson and rank correlation | About 0.97 and about 0.95 | Diener et al., PLCMOS paper 2023 |
| DNSMOS (a general noise-suppression metric): file-wise vs model-wise Pearson correlation | 0.52 vs 0.85 | Diener et al., PLCMOS paper 2023 |
| PESQ vs PLCMOS on reference-aligned data: model-wise Pearson correlation | 0.52 vs 0.98 | Diener et al., PLCMOS paper 2023 |
| STOI (an intelligibility metric): model-wise Pearson correlation | 0.21 | Diener et al., PLCMOS paper 2023 |
| Human votes in the main PLCMOS training set (LibriSpeech Basic) | 333,740 | Diener et al., PLCMOS paper 2023 |
| Test files in the IETF machine-learning codec listening test battery | 100 clean speech, 90 noise or reverberation, 20 overlapping talkers | IETF draft-ietf-mlcodec-test-battery-00, June 2026 |
Context: crowdsourced ratings correlated with expert ratings at about 0.83 PCC (95% CI: [0.78, 0.87]) in the PLCMOS labeling pre-study. The IETF test battery draft still lists its packet loss scenarios as “Details TBD”, so there is no standard loss test for neural codecs yet.
6. Redundancy Coding: Paying Bits to Avoid Concealment
Concealment guesses; redundancy sends a backup. Forward error correction (FEC) means attaching a lower-quality copy of earlier audio to later packets, so a lost packet can be rebuilt instead of invented. The trade-off is bandwidth. Opus’s classic in-band FEC (LBRR) covers just one frame at about 2/3 of the regular bitrate, while Deep REDundancy (DRED) carries up to one second or more at about 1/50 of the regular bitrate (IETF draft-ietf-mlcodec-opus-dred-07, August 2026).
The older RED approach shows the scale of the effect. In a Chrome test at 20% loss, sending one previous frame in each packet cut the concealed share of audio from 18% to 4%, and two previous frames cut it to 0.7% (webrtcHacks, RED: Improving Audio Quality with Redundancy 2020). The cost: audio rose from about 30 kbps to 60 kbps with one copy and 90 kbps with two. DRED exists to get similar protection without doubling the bitrate.
| Metric | Value | Source |
|---|---|---|
| Opus LBRR in-band FEC: depth of redundancy and bitrate cost | 1 frame; about 2/3 of the regular Opus bitrate | IETF draft-ietf-mlcodec-opus-dred-07, 2026 |
| Opus DRED: depth of redundancy and bitrate cost | Up to one second or more per packet; about 1/50 of the regular bitrate | IETF draft-ietf-mlcodec-opus-dred-07, 2026 |
| DRED evaluation: redundancy carried per packet, and number of times each frame is sent | 1.04 seconds; 52 copies | Valin et al., DRED paper (arXiv 2212.04453 v3, 2024) |
| Bitrate of the LBRR test condition in the DRED evaluation (24 kb/s base plus LBRR) | 40 kb/s total, LBRR averaging about 16 kb/s | Valin et al., DRED paper 2024 |
| Concealed share of audio at 20% loss in Chrome: no RED vs one previous frame vs two previous frames | 18% vs 4% vs 0.7% | webrtcHacks, Hancke RED test 2020 |
| Concealed share of audio at 60% random loss: no RED vs one vs two previous frames | 60% vs 32% vs 18% | webrtcHacks, Hancke RED test 2020 |
| Minimum Opus bitrate to encode in-band FEC at 10% packet loss | 19 kbps (vendor’s own comparison) | Meta Engineering, MLow codec blog 2024 |
| Bitrate of Meta’s MLow samples under 30% receiver-side packet loss | 14 kbps (vendor’s own claim) | Meta Engineering, MLow codec blog 2024 |
Context: the webrtcHacks figures are a single-author lab test from 2020 (most recent available public measurement of RED in Chrome). Meta’s figures come from its own MLow announcement and have not been independently replicated. All DRED differences in the paper’s P.808 listening test (966 utterances, 10 listeners each) were reported as significant at p < .05.
7. Shipping Concealment: Compute, Footprint and Standards
A concealment model only matters if it fits on the cheapest phone in the call. Opus 1.5’s Deep PLC costs about 1% of a laptop CPU core at high loss rates and adds about 1 MB to the library (Xiph.Org, Opus 1.5 demo, March 2024, the developers’ own figures). The 2022 challenge entry from the same team needed 13% to 14% of a core (section 3), although the two figures were measured differently and are not a strict like-for-like comparison.
The standards layer, by contrast, barely moves. WebRTC’s audio rules (RFC 7874, 2016) mandate Opus and G.711 but set no requirement for concealment or FEC, leaving quality under loss entirely to each implementation. Opus 1.6 continued shrinking the models: its DRED networks are about 3x smaller than in 1.5 (Xiph.Org, Opus 1.6 demo, December 2025).
| Metric | Value | Source |
|---|---|---|
| Opus 1.5 Deep PLC: added library size and CPU cost | About 1 MB; about 1% of a laptop core at high loss; needs decoder complexity 5 or higher | Xiph.Org, Opus 1.5 demo 2024 (developers’ own figures) |
| Opus 1.5 DRED: added library size, run-time cost and bitrate overhead | About 2 MB; about 1%; about 12 to 32 kb/s | Xiph.Org, Opus 1.5 demo 2024 (developers’ own figures) |
| FARGAN neural vocoder used to synthesize concealed and redundant speech: complexity | 600 MFLOPS, 1/5 of LPCNet | Xiph.Org, Opus 1.5 demo 2024 |
| Loss rates in the Opus 1.5 WebRTC tests (PLCMOS v2) | 0% to 90% in 10% steps; DRED reported to hold quality at 90% loss | Xiph.Org, Opus 1.5 demo 2024 (developers’ own claim) |
| Opus 1.6 DRED model size vs Opus 1.5 | 600 kB vs 1,800 kB | Xiph.Org, Opus 1.6 demo, December 2025 |
| Opus frame sizes, packet length and bitrate range | 2.5 to 60 ms frames; packets up to 120 ms; 6 to 510 kbit/s | IETF RFC 6716 (2012) |
| Opus algorithmic delay | 5 ms to 65.2 ms | IETF RFC 6716 (2012) |
| Codecs WebRTC endpoints must support | Opus and G.711 (PCMA and PCMU); no PLC or FEC requirement | IETF RFC 7874 (2016) |
Context: RFC 6716 and RFC 7874 are older than three years but remain the current standards (RFC 6716, RFC 7874). The DRED draft, if approved, would update RFC 6716 with a neural redundancy layer.
Summary: WebRTC Audio Packet Loss Concealment by the Numbers
| Metric | Value | Source |
|---|---|---|
| Google Duo calls that must deal with packet loss, jitter or delay | 99% | Google Research, WaveNetEQ blog 2020 |
| Duo calls that lose more than 8% of their audio | 10% | Google Research, WaveNetEQ blog 2020 |
| Teams Poor threshold for average audio packet loss | Above 0.1 (10%) | Microsoft Learn, CQD Stream Classification 2026 |
| Twilio high_packet_loss tag threshold | Above 5% | Twilio Voice Insights, Details: Call Tags |
| G.711 robustness factor Bpl: no concealment vs Appendix I concealment | 4.3 vs 25.1 | ITU-T G.113 (09/2024) |
| G.729 Annex A with voice activity detection: Bpl | 19.0 | ITU-T G.113 (09/2024) |
| 2022 challenge PLCMOS: zero-fill vs winner vs clean | 2.90 vs 4.28 vs 4.51 | Microsoft, INTERSPEECH 2022 Deep PLC Challenge paper |
| 2022 challenge word accuracy: zero-fill vs winner vs clean | 0.86 vs 0.88 vs 0.97 | Microsoft, INTERSPEECH 2022 Deep PLC Challenge paper |
| Opus team’s hybrid PLC compute per 10 ms of speech | 1.3 to 1.4 ms (13% to 14% of one core) | Valin et al. 2022 |
| 2024 challenge P.804 overall quality: lossy vs best vs clean | 2.37 vs 3.49 vs 4.01 | Microsoft, ICASSP 2024 Deep PLC Challenge paper |
| 2024 challenge systems scoring below unprocessed lossy audio | 2 of 8 | Microsoft, ICASSP 2024 Deep PLC Challenge paper |
| PLCMOS model-wise Pearson correlation with human ratings | About 0.97 | Diener et al., PLCMOS paper 2023 |
| PESQ vs STOI model-wise Pearson correlation on PLC systems | 0.52 vs 0.21 | Diener et al., PLCMOS paper 2023 |
| LBRR vs DRED redundancy depth | 1 frame vs up to one second or more | IETF draft-ietf-mlcodec-opus-dred-07, 2026 |
| DRED redundancy bitrate relative to regular Opus | About 1/50 | IETF draft-ietf-mlcodec-opus-dred-07, 2026 |
| Concealed audio at 20% loss: no RED vs two-frame RED | 18% vs 0.7% | webrtcHacks, Hancke RED test 2020 |
| Minimum Opus in-band FEC bitrate at 10% loss | 19 kbps (vendor’s own comparison) | Meta Engineering, MLow codec blog 2024 |
| Opus 1.5 Deep PLC CPU cost | About 1% of a laptop core | Xiph.Org, Opus 1.5 demo 2024 |
| Opus 1.6 DRED model size vs 1.5 | 600 kB vs 1,800 kB | Xiph.Org, Opus 1.6 demo 2025 |
Methodology and Sources
- Microsoft Research, INTERSPEECH 2022 Audio Deep Packet Loss Concealment Challenge paper (Diener et al., 2022): arxiv.org/pdf/2204.05222
- Microsoft Research, ICASSP 2024 Audio Deep Packet Loss Concealment Challenge paper (2024): arxiv.org/html/2402.16927v1
- Microsoft Research, ICASSP 2024 Audio Deep PLC Challenge program page: microsoft.com
- Diener et al., PLCMOS: a data-driven non-intrusive metric for the evaluation of packet loss concealment algorithms (Interspeech 2023): arxiv.org/pdf/2305.15127
- Valin et al., Real-Time Packet Loss Concealment With Mixed Generative and Predictive Model (2022): arxiv.org/pdf/2205.05785
- Valin et al., DRED: Deep REDundancy Coding of Speech Using a Rate-Distortion-Optimized Variational Autoencoder (arXiv v3, 2024): arxiv.org/pdf/2212.04453
- IETF, Deep Audio Redundancy (DRED) Extension for the Opus Codec, draft-ietf-mlcodec-opus-dred-07 (August 2026): ietf.org
- IETF, ML codec listening test battery, draft-ietf-mlcodec-test-battery-00 (June 2026): ietf.org
- IETF RFC 6716, Definition of the Opus Audio Codec (2012): rfc-editor.org
- IETF RFC 7874, WebRTC Audio Codec and Processing Requirements (2016): rfc-editor.org
- W3C, Identifiers for WebRTC’s Statistics API (Candidate Recommendation Draft, 25 September 2025): w3.org
- ITU-T G.113 (09/2024), Transmission impairments due to speech processing: itu.int
- ITU-T G.107 (06/2015), The E-model: itu.int
- Google Research, Improving Audio Quality in Duo with WaveNetEQ (April 2020): research.google
- Microsoft Learn, Stream Classification in Call Quality Dashboard (updated June 2026): learn.microsoft.com
- Microsoft Learn, Dimensions and measures available in Call Quality Dashboard (updated July 2026): learn.microsoft.com
- Twilio, Voice Insights Details: Call Tags: twilio.com
- Meta Engineering, MLow: Meta’s low bitrate audio codec (June 2024): engineering.fb.com
- Xiph.Org, Opus 1.5 demo (March 2024) and Opus 1.6 demo (December 2025): opus-codec.org/demo/opus-1.5, opus-codec.org/demo/opus-1.6
- webrtcHacks, RED: Improving Audio Quality with Redundancy (Philipp Hancke, August 2020): webrtchacks.com
- Data watch: the only call-level loss distribution from a major platform is Google’s 2020 Duo figure, and the RED concealment measurements are a 2020 single-author test, so both are flagged as most recent available data. The E-model example in section 2 is our own calculation from ITU-T G.107 Equation 7-29 and Annex B with G.113 Bpl values, holding every other input at default; it is an estimate, not a measurement. Challenge scores are not comparable across editions: 2022 used wideband audio, PLCMOS and CMOS with bursts up to 1 second, while 2024 used full-band audio, P.804 and bursts up to 3 seconds. In the 2022 paper, PLCMOS correlated with CMOS at about 0.95 Pearson, while the later PLCMOS release reports about 0.97 on a different evaluation set. Xiph.Org and Meta figures are the developers’ own claims. Thresholds from Teams (10%) and Twilio (5%) conflict by design because each vendor tunes them to its own media stack. The DRED draft (expires February 2027) and the IETF test battery draft, whose packet loss scenarios are still undefined, may change before publication as RFCs.
Last updated: October 24, 2026. We update this roundup quarterly, and the next refresh is expected when the IETF mlcodec working group publishes the next revision of the Opus DRED draft and the test battery’s packet loss scenarios.