The 2021 Deep Noise Suppression (DNS) Challenges capped total algorithmic latency at 40 ms, while a 2025 hearables study ran neural noise suppression at 1.25 ms of algorithmic latency and 3.35 ms end to end on a low-power DSP (Microsoft Research, Interspeech 2021 DNS Challenge paper; Dementyev et al., arXiv 2025). Model sizes span more than three orders of magnitude: RNNoise uses 87,503 weights and GTCRN reports 23.7K parameters, while the top Interspeech 2025 URGENT system used about 102M. Listening tests keep exposing what lab metrics miss: the full PercepNet had a lower PESQ than its variant without a post-filter (2.54 vs 2.73) but a higher listener MOS (4.05 vs 3.91). Noise suppression is half of the voice isolation story; the other half, separating one talker from other talkers, is covered in our speech separation and cocktail party statistics, and the business side of these tools is in our enterprise noise cancelling software statistics. We aggregated data from Microsoft Research’s DNS Challenge papers, ITU-T, Mozilla and Xiph.Org’s RNNoise work, the DeepFilterNet papers, the URGENT challenge organizers, NVIDIA’s Maxine documentation, and independent university and industry researchers.
TL;DR
- A 1.25 ms neural suppressor ran on a low-power DSP with 3.35 ms end-to-end latency (Dementyev et al., sub-millisecond hearables paper, arXiv 2025).
- RNNoise runs on 87,503 weights, about 1.3% of one x86 core (Mozilla, RNNoise paper, MMSP 2018).
- GTCRN reports 23.7K parameters and 39.6 MMACs per second (Rong et al., GTCRN, ICASSP 2024).
- DeepFilterNet2 has 2.306 M parameters and a real-time factor of 0.42 even on a Raspberry Pi 4 (Schröter et al., DeepFilterNet2 paper 2022).
- A GPU denoiser needs 0.410 ms (p95) to process a 10 ms frame on a laptop GPU (NVIDIA, Maxine AFX SDK performance guide, vendor claim).
- PercepNet scored a listener MOS of 4.05 against 3.40 for the noisy input on VCTK (Valin et al., PercepNet paper, Interspeech 2020).
- The Interspeech 2025 URGENT winner used about 102M parameters, about 12 times the 8.5M baseline (Saijo et al., URGENT 2025 Challenge paper).
- Predicting ahead to reach 0 ms latency dropped SI-SDR from 2.8341 dB to -3.6493 dB (Wang et al., IEEE/ACM TASLP 2022).
- ITU-T P.835, the three-rating test for noise suppressors, has a new edition, P.835 (07/26) (ITU-T Recommendation P.835 page, 2026).
- The ICASSP 2026 URGENT Challenge drew 80 Track 1 registrations and valid entries from 29 teams (Li et al., ICASSP 2026 URGENT Challenge paper).
1. Latency Budgets: How Much Delay Real-Time Noise Suppression Gets
Algorithmic latency is the delay built into the method itself (the length of the analysis frame plus any look-ahead into future audio), so a faster CPU cannot remove it. Processing time is separate: a real-time system only has to finish each frame before the next one arrives. The 2021 DNS Challenges capped total algorithmic latency at 40 ms, which is generous by hearing-aid standards and tight for a model that wants to see future audio.
That 40 ms is spent before the audio even leaves the computer. On a call, the suppressor’s delay stacks with echo cancellation, codecs and the network; echo control has its own budget and benchmarks, covered in our acoustic echo cancellation statistics. The rule is spelled out in the Interspeech 2021 DNS Challenge paper, and the design figures come from the PercepNet paper and the DeepFilterNet3 paper.
| Metric | Value | Source |
|---|---|---|
| DNS Challenge real-time rule: total algorithmic latency (frame + stride + look-ahead), and processing time per frame | 40 ms or less; under the stride time on an Intel Core i5 quad-core at 2.4 GHz (a 32 ms frame with a 16 ms stride, 48 ms, fails) | Microsoft Research, ICASSP 2021 and Interspeech 2021 DNS Challenge papers |
| RNNoise analysis window, hop and look-ahead | 20 ms windows, 10 ms offset; 10 ms look-ahead | Mozilla, RNNoise paper, MMSP 2018; Xiph.Org RNNoise demo page 2017 |
| PercepNet frame size and look-ahead | 10 ms frames, 40 ms look-ahead | Valin et al., PercepNet paper, Interspeech 2020 |
| DeepFilterNet3 algorithmic latency | 40 ms (20 ms window, 10 ms hop, 2 frames of look-ahead) | Schröter et al., DeepFilterNet3 paper, Interspeech 2023 |
| Very-low-latency STFT design: window sizes and resulting algorithmic latency | 16 ms input window, 4 ms output window, 2 ms hop; 4 ms latency, or 2 ms with one-frame-ahead prediction | Wang et al. (CMU, MERL), IEEE/ACM TASLP 2022 |
| Typical ML speech enhancement algorithmic latency vs hearables target | 16 to 32 ms vs under 2 ms end to end (about 1 ms or less algorithmic) | Dementyev et al., sub-millisecond hearables paper, arXiv 2025 |
| Denoiser v2 GPU effect: frame size and per-frame release threshold (vendor’s own test) | 10 ms frames (160 samples at 16 kHz); p95 frame time must stay below the 10 ms budget | NVIDIA, Maxine AFX SDK Windows performance guide (vendor claim) |
Most recent available data for the challenge rule: 2021; later DNS editions defer their millisecond definitions to the challenge website. Most recent available data for RNNoise and PercepNet: 2018 and 2020 papers.
2. Compute and Model Size: From 23.7K Parameters to 8 Million
GTCRN reports 23.7K parameters, while PercepNet uses 8M weights and about 800 MMACS, roughly 340 times more weights (8,000,000 / 23,700) for a model that also runs in real time. Real-time factor (RTF) is processing time divided by audio duration, so an RTF below 1 means the model keeps up with live audio.
The practical reading: compute is no longer what stops neural suppression on a laptop. RNNoise needed about 1.3% of one core on a Haswell-era laptop CPU and PercepNet less than 5% of a core. The open question is phones, earbuds and DSPs, where a 0.42 RTF on a Raspberry Pi 4 still leaves little headroom. Figures come from the RNNoise paper, the DeepFilterNet2 paper and the GTCRN paper page.
| Metric | Value | Source |
|---|---|---|
| RNNoise size and CPU load | 87,503 weights; about 40 MFLOPS; about 1.3% of one Intel i7-4800MQ core, about 14% of one Raspberry Pi 3 core | Mozilla, RNNoise paper, MMSP 2018 |
| RNNoise model file with 8-bit weights vs 32-bit floats, and speed | 85 kB vs 340 kB; about 60x faster than real time on x86, about 7x on a Raspberry Pi 3 | Xiph.Org, RNNoise demo page 2017 |
| PercepNet size and CPU load | 8M weights; about 800 MMACS; 5.2% of one Intel i7-8565U core (10 ms frames), 4.1% (40 ms frames) | Valin et al., PercepNet paper, Interspeech 2020 |
| DeepFilterNet vs DeepFilterNet2: parameters, compute, RTF on a Core i5-8250U | 1.778 M vs 2.306 M; 0.348 vs 0.356 GMACs; RTF 0.11 vs 0.04 (0.42 on a Raspberry Pi 4) | Schröter et al., DeepFilterNet2 paper 2022 |
| GTCRN size and compute | 23.7K parameters; 39.6 MMACs per second | Rong et al., GTCRN, ICASSP 2024 |
| UL-UNAS size and compute | 169k parameters; 34M MACs per second; 0 ms look-ahead | Rong et al., UL-UNAS, Neural Networks 2025 |
| Denoiser v2 on an RTX 3500 Ada laptop GPU (vendor’s own measurement) | p95 0.410 ms per 10 ms frame; p50 throughput 26.74x real time; 215 MiB GPU memory | NVIDIA, Maxine AFX SDK Windows performance guide (vendor claim) |
NVIDIA’s denoiser documentation adds its own caveat: input with an SNR below 5 dB may produce distorted output.
3. Benchmark Quality Scores: PESQ and DNSMOS for Named Models
PESQ compares the processed audio with a clean reference; DNSMOS is a neural model that predicts what listeners would say without needing a reference. On the common VoiceBank+DEMAND test set, PESQ climbed from 1.97 for unprocessed noisy speech to 3.17 for DeepFilterNet3, with RNNoise at the low end of the neural pack.
The DNSMOS row holds the more useful warning. On the DNS4 blind test set, RNNoise lowered the predicted speech (SIG) score below the noisy input while raising the overall score, which is the classic trade: less noise, slightly damaged voice. PercepNet’s own paper shows the reverse problem for PESQ, where the version listeners liked most had a lower PESQ. Data from the DeepFilterNet3 paper and the UL-UNAS paper.
| Metric | Value | Source |
|---|---|---|
| VoiceBank+DEMAND PESQ: noisy input vs RNNoise vs DeepFilterNet2 (with post-filter) | 1.97 vs 2.33 vs 3.03 | Schröter et al., DeepFilterNet2 paper 2022 |
| VoiceBank+DEMAND PESQ and STOI of DeepFilterNet3 | 3.17 and 0.944 | Schröter et al., DeepFilterNet3 paper, Interspeech 2023 |
| DNSMOS on the DNS4 blind set, speech (SIG) and overall (OVRL): noisy vs RNNoise vs DeepFilterNet2 | SIG 4.144 vs 3.884 vs 4.196; OVRL 3.291 vs 3.378 vs 3.882 | Schröter et al., DeepFilterNet2 paper 2022 |
| VCTK-DEMAND PESQ: GTCRN vs UL-UNAS trained with a PESQ loss | 2.87 vs 3.09 | Rong et al., UL-UNAS, Neural Networks 2025 |
| DNS3 blind set PESQ and look-ahead: GTCRN vs LiSenNet vs UL-UNAS | 2.10 (0 ms) vs 2.23 (32 ms) vs 2.28 (0 ms) | Rong et al., UL-UNAS, Neural Networks 2025 |
| PercepNet PESQ-WB with vs without its post-filter | 2.54 vs 2.73 | Valin et al., PercepNet paper, Interspeech 2020 |
| P.808 listener MOS on VCTK 48 kHz: noisy vs RNNoise vs PercepNet | 3.40 vs 3.70 vs 4.05 | Valin et al., PercepNet paper, Interspeech 2020 |
Context: VoiceBank+DEMAND is a small, mostly synthetic test set, so these rankings say more about relative progress than about real calls. The DNS3 row is the one latency readers should note: a zero look-ahead model (2.28) scored above a 32 ms look-ahead model (2.23) on PESQ.
4. Challenge Results: DNS 2021 to 2023 and URGENT 2025 to 2026
Challenge results from 2020 to 2026 show that measured gains over the noisy input stay small, even for winners. At ICASSP 2021 the best model was only 0.53 DMOS (difference in mean opinion score) better than the noisy input, which the organizers put at an absolute MOS of 3.38, and only half of the submissions beat the baseline. Personalized suppression, which Teams ships as voice isolation, enrolls a sample of the user’s voice and also removes other talkers; in the ICASSP 2022 personalized track no team improved the speech score. In Teams, voice isolation is enabled by default in admin policy and the voice profile is stored on the user’s device (Microsoft Learn, Teams voice isolation admin page, updated 2026).
URGENT (“universal” speech enhancement) is broader than noise suppression: it adds reverberation, clipping, codec loss, packet loss and wind noise, and takes 8 to 48 kHz input. Its 2025 results show model size moving in the opposite direction from section 2. Sources: ICASSP 2021 DNS paper, ICASSP 2022 DNS paper, ICASSP 2023 DNS paper, URGENT 2025 paper and URGENT 2026 paper.
| Metric | Value | Source |
|---|---|---|
| Interspeech 2020 DNS real-time track: PercepNet rank and P.808 MOS vs noisy and baseline | 2nd place; 3.52 vs 3.01 vs 3.03 | Valin et al., PercepNet paper, Interspeech 2020 |
| ICASSP 2021 DNS Track 1: teams and best result vs noisy input | 16 teams; best model 0.53 DMOS better than noisy (absolute MOS 3.38); half of submissions beat the baseline | Microsoft Research, ICASSP 2021 DNS Challenge paper |
| Interspeech 2021 DNS: best overall MOS vs ceiling with perfect noise removal; speech scores | About 3.78 vs 3.92; most teams other than the top two scored below noisy on speech MOS | Microsoft Research, Interspeech 2021 DNS Challenge paper |
| ICASSP 2022 personalized (voice isolation) track: enrollment length and result vs noisy | 2.5 minutes per talker; no team improved SIG; word accuracy 2% worse | Microsoft Research, ICASSP 2022 DNS Challenge paper |
| ICASSP 2023 DNS: submissions and winning Score gain over the noisy blind set, headset vs speakerphone | 11 and 11 submissions; 0.145 vs 0.141 | Microsoft Research, ICASSP 2023 DNS Challenge paper |
| Interspeech 2025 URGENT: submissions and size of the winning system vs baseline | 32 submissions; about 102M vs about 8.5M parameters | Saijo et al., Interspeech 2025 URGENT Challenge paper |
| Interspeech 2025 URGENT: purely generative system’s MOS, overall rank and character accuracy on Chinese and Japanese | Highest MOS (3.69) but ranked 13th; 20.1% and 36.8% | Saijo et al., Interspeech 2025 URGENT Challenge paper |
| ICASSP 2026 URGENT: Track 1 registrations, valid teams, blind test size | 80 registrations; 29 teams with valid entries across both tracks; 840 samples including 5 unseen languages | Li et al., ICASSP 2026 URGENT Challenge paper |
The ceiling of 3.92 in the Interspeech 2021 row is the organizers’ own calculation: if noise were rated a perfect 5 and speech stayed as good as in the noisy clip, overall MOS would top out there. By 2021 the winners had removed most of the noise and the remaining gap was speech quality. The ICASSP 2023 Score is 0.5 x [WAcc + 0.25 x (OVRL - 1)], so it weights word accuracy and listener quality equally.
5. Measuring Quality: ITU-T P.835, P.808 and DNSMOS
ITU-T P.835 asks listeners for three separate ratings: the speech signal (SIG), the background noise (BAK) and overall quality (OVRL). That split matters for noise suppression because a model can win on noise and lose on voice in the same clip. DNSMOS tracked human ratings at a PCC of 0.93 per model, against 0.78 for PESQ and just 0.23 for SDR, which is why the DNS organizers trained a metric on listener ratings instead of relying on PESQ.
The catch is granularity. The P.835 version of DNSMOS is far weaker on a single clip than when averaged over a whole model, so treat a single-clip DNSMOS number as a hint, not a verdict. The same logic applies to synthetic voices, covered in our synthetic speech naturalness MOS statistics. Sources: ITU-T P.835 page, ITU-T P.808 page, DNSMOS paper and DNSMOS P.835 paper.
| Metric | Value | Source |
|---|---|---|
| ITU-T P.835 editions | P.835 (07/26) in force, prepublished; P.835 (11/03) superseded | ITU-T Recommendation P.835 page, accessed October 2026 |
| ITU-T P.808 editions | P.808 (06/21) in force; P.808 (06/18) superseded | ITU-T Recommendation P.808 page, accessed October 2026 |
| DNSMOS correlation with listeners per model (PCC) vs PESQ, POLQA, SDR; DNSMOS P.835 overall, model level vs single clip | 0.93 vs 0.78, 0.79, 0.23; 0.98 vs 0.82 | Reddy et al., Microsoft, DNSMOS (2020) and DNSMOS P.835 (2021) papers |
| PercepNet listening tests: VCTK samples and listeners per sample; DNS test listeners per clip | 824 samples, 8 listeners each (95% CI 0.04); 10 listeners each (95% CI 0.02) | Valin et al., PercepNet paper, Interspeech 2020 |
| URGENT 2025 subjective test method and scope | ITU-T P.808 absolute category rating on Amazon Mechanical Turk; English data only | Saijo et al., Interspeech 2025 URGENT Challenge paper |
| ICASSP 2026 URGENT listening tests and MOS-prediction data | Top 6 systems rated with P.808 ACR and CCR; 8,000 MOS-annotated samples from 16 systems | Li et al., ICASSP 2026 URGENT Challenge paper |
The P.835 revision is the newest item here: the 2003 edition was superseded by P.835 (07/26), and its title now covers terminals “providing noise suppression functionality”. Its full text was not yet public when we checked, so changes to the method are not reflected in the numbers above.
6. The Latency vs Quality Trade-off
Cutting latency costs quality and compute at the same time. In a 2025 hearables study, a 1 ms model needed 840 MIPS while a 16 ms overlap-add model needed 105 MIPS, and the 16 ms model still scored higher (DNSMOS OVRL 2.35 vs 2.13 on CHiME-2). An earlier CMU and MERL study found the same shape for PESQ, and showed that trying to predict ahead of the audio to reach 0 ms collapses SI-SDR.
The hearables study’s answer was to change the output filter rather than shrink the window: a minimum-phase FIR design held SI-SDR roughly flat from 0.32 to 1.25 ms and cut end-to-end latency from 7.1 ms to 3.35 ms. Live chat is far more tolerant of delay than hearables, but the suppressor’s latency still adds to every other stage, for example when effects are stacked in a Discord voice changer chain. Data from the hearables paper and the very-low-latency STFT paper.
| Metric | Value | Source |
|---|---|---|
| DNSMOS OVRL on CHiME-2: unprocessed vs 1 ms vs 2 ms vs 4 ms vs 16 ms model | 1.3 vs 2.13 vs 2.25 vs 2.28 vs 2.35 | Dementyev et al., sub-millisecond hearables paper, arXiv 2025 |
| SI-SDR on CHiME-2: 1 ms vs 2 ms vs 16 ms model | 8.0 dB vs 11.10 dB vs 12.98 dB | Dementyev et al., sub-millisecond hearables paper, arXiv 2025 |
| Compute of the 1 ms model vs the 16 ms overlap-add model | 840 MIPS vs 105 MIPS | Dementyev et al., sub-millisecond hearables paper, arXiv 2025 |
| Proposed 1.25 ms Deep FIR model: parameters, compute, end-to-end latency on an NXP i.MX RT600 | 626k parameters; 376 MIPS; 3.35 ms | Dementyev et al., sub-millisecond hearables paper, arXiv 2025 |
| Single-microphone PESQ, rectangular-window system: unprocessed vs 4 ms vs 2 ms vs 0 ms (predicting ahead) | 1.44 vs 1.8997 vs 1.84575 vs 1.7138 | Wang et al. (CMU, MERL), IEEE/ACM TASLP 2022 |
| Single-microphone SI-SDR, rectangular-window system, at 4 ms vs 0 ms (predicting ahead) | 2.8341 dB vs -3.6493 dB | Wang et al. (CMU, MERL), IEEE/ACM TASLP 2022 |
| Six-microphone PESQ at 4 ms vs 2 ms (same two-stage system) | 2.698490 vs 2.47482 | Wang et al. (CMU, MERL), IEEE/ACM TASLP 2022 |
| DeepFilterNet window-size test: where complex ratio masks start losing quality | FFT windows of 20 ms or less (deep filtering stays stable from 5 to 30 ms) | Schröter et al., DeepFilterNet paper, ICASSP 2022 |
Outlier note: the 1 ms and shorter models in the hearables study could not run in real time on the target DSP, so their end-to-end figures are the authors’ estimates.
Summary: Neural Noise Suppression and AI Voice Isolation Latency by the Numbers
| Metric | Value | Source |
|---|---|---|
| Deep FIR hearables model running in real time on a low-power DSP | 1.25 ms algorithmic, 3.35 ms end to end | Dementyev et al., arXiv 2025 |
| ICASSP 2026 URGENT blind test size | 840 samples, including 5 unseen languages | Li et al., ICASSP 2026 URGENT Challenge paper |
| RNNoise weights | 87,503 | Mozilla, RNNoise paper, MMSP 2018 |
| GTCRN parameters and compute | 23.7K; 39.6 MMACs per second | Rong et al., GTCRN, ICASSP 2024 |
| UL-UNAS parameters and look-ahead | 169k; 0 ms | Rong et al., UL-UNAS, Neural Networks 2025 |
| VoiceBank+DEMAND PESQ, noisy vs DeepFilterNet3 | 1.97 vs 3.17 | Schröter et al., DeepFilterNet2 and DeepFilterNet3 papers |
| Denoiser v2 p95 time per 10 ms frame on a laptop GPU (vendor claim) | 0.410 ms | NVIDIA, Maxine AFX SDK performance guide |
| PercepNet listener MOS vs noisy on VCTK | 4.05 vs 3.40 | Valin et al., PercepNet paper, Interspeech 2020 |
| URGENT 2025 winner vs baseline size | About 102M vs 8.5M parameters | Saijo et al., URGENT 2025 Challenge paper |
| URGENT 2025 generative system: listener MOS vs overall rank | 3.69 (highest) vs 13th | Saijo et al., URGENT 2025 Challenge paper |
| ICASSP 2026 URGENT valid teams | 29 | Li et al., ICASSP 2026 URGENT Challenge paper |
| Single-microphone PESQ at 4 ms vs 0 ms latency | 1.8997 vs 1.7138 | Wang et al., IEEE/ACM TASLP 2022 |
| Newest ITU-T P.835 edition | P.835 (07/26), in force (prepublished) | ITU-T Recommendation P.835 page |
| Edition of ITU-T P.808 in force | P.808 (06/21) | ITU-T Recommendation P.808 page |
| RNNoise model size with 8-bit weights | 85 kB | Xiph.Org, RNNoise demo page 2017 |
| RNNoise speed on x86 | About 60x faster than real time | Xiph.Org, RNNoise demo page 2017 |
Methodology and Sources
- Microsoft (Research papers and Learn documentation): ICASSP 2021 DNS Challenge paper (2020), Interspeech 2021 DNS Challenge paper (2021), ICASSP 2022 DNS Challenge paper (2022), ICASSP 2023 DNS Challenge paper (2023), Reddy et al. DNSMOS (2020) and DNSMOS P.835 (2021), and the Teams voice isolation admin page (updated 2026).
- ITU-T, Recommendation P.835 and Recommendation P.808, accessed October 2026.
- Mozilla, RNNoise paper (MMSP 2018) and Xiph.Org RNNoise demo page (2017).
- Valin et al., PercepNet paper (Interspeech 2020).
- Schröter et al., DeepFilterNet (ICASSP 2022), DeepFilterNet2 (2022) and DeepFilterNet3 (Interspeech 2023).
- Rong et al., GTCRN (ICASSP 2024) and UL-UNAS (Neural Networks 2025).
- Wang et al. (CMU, MERL), STFT-Domain Neural Speech Enhancement with Very Low Algorithmic Latency (IEEE/ACM TASLP 2022).
- Dementyev et al., Towards Sub-millisecond Latency Real-Time Speech Enhancement Models on Hearables (arXiv v2, March 2025).
- Saijo et al., Interspeech 2025 URGENT Challenge (2025) and Li et al., ICASSP 2026 URGENT Challenge (2026).
- NVIDIA, Maxine AFX SDK performance guide and denoiser effect page (vendor documentation, labelled as vendor claims).
- Data watch: most DNS Challenge data is 2020 to 2023; the ICASSP 2023 paper is the latest DNS Challenge report we found, so it is the most recent available data for that benchmark series. RNNoise PESQ on VoiceBank+DEMAND is 2.33 in the DeepFilterNet2 table but 2.29 in the PercepNet and UL-UNAS tables. GTCRN appears as 23.7K parameters and 39.6 MMACs in the ICASSP 2024 abstract, but the UL-UNAS paper quotes updated repository figures of 48k parameters and 34 MMACs. DeepFilterNet2 reports an RTF of 0.04 and DeepFilterNet3 an RTF of 0.19 (single-threaded) on the same Core i5-8250U, so threading or measurement settings differ. The hearables paper gives 0.32 ms in its abstract and 0.34 ms in its introduction for the shortest latency. Values from the CMU and MERL low-latency paper are quoted at the precision printed in its tables. The vendor GPU figures come from one laptop GPU and are not guarantees. Per-team SIG, BAK and OVRL values for the 2022 and 2023 DNS editions exist only as charts, so we did not transcribe them. The full text of ITU-T P.835 (07/26) was not yet public.
Last updated: October 25, 2026. We update this roundup quarterly, and the next refresh is expected when the ICASSP 2026 URGENT Challenge proceedings and the published text of ITU-T P.835 (07/26) become available.