AI voice isolation cut speech-to-text word error rate by 73% across 265 real conversations in a September 2026 benchmark, but slightly increased errors on recordings that were already clean (Krisp, Voice Isolation Benchmark 2026). That trade-off frames the whole category: noise suppression is now a default setting in mainstream meeting tools, yet its value depends on how dirty the audio is to begin with. Google says Meet noise cancellation is on by default for Business and Enterprise accounts, Microsoft lists Teams voice isolation as enabled by default for users, and Owl Labs found 76% of US workers cite audio echo or distortion as a common hybrid meeting problem (Owl Labs, State of Hybrid Work 2025). The broader collaboration stack these features live in is covered in our enterprise unified communications statistics. We aggregated data from Microsoft Research’s Deep Noise Suppression Challenge papers, Microsoft, Google, Zoom, Apple, NVIDIA, Krisp, Owl Labs, NIOSH, the World Health Organization and peer-reviewed audiology and speech research indexed in Europe PMC.
TL;DR
- Voice isolation reduced word error rate by 73% overall across 265 recorded conversations (Krisp, Voice Isolation Benchmark 2026).
- Work-environment WER fell from 31.84% to 6.30%, but clean phone calls went from 3.48% to 3.91% (Krisp, Voice Isolation Benchmark 2026).
- Google Meet noise cancellation is on by default for Business and Enterprise accounts (Google, Meet Help Center 2026).
- The Teams VoiceIsolation policy value is Enabled by default (Microsoft Learn, voice isolation admin guide, April 2026).
- Teams had more than 320 million monthly active users (Microsoft, FY2024 Q1 earnings call, October 2023).
- 76% of US workers cite audio echo or distortion and 75% not hearing everyone as common hybrid meeting issues (Owl Labs, State of Hybrid Work 2025).
- About 27 million US workers are exposed to hazardous noise each year (NIOSH, CDC).
- The first DNS Challenge released 441 hours of clean speech and drew 28 submissions from 19 teams (Reddy et al., Interspeech 2020).
- By 2023 the challenge required a combined latency of 20 ms or less and used 760.53 hours of clean speech in one track (Dubey et al., ICASSP 2023 DNS Challenge).
- 60% of CX leaders plan to adopt AI voice translation within 6 to 12 months, up from 36% in 2025 (Krisp, 2026 State of Voice in CX).
- A DNN hearing aid program raised sentence recognition in noise to 79% from 60% for bimodal cochlear implant users (Kolberg et al., J Clin Med 2025).
- 430 million people need rehabilitation for disabling hearing loss (WHO, Deafness and hearing loss fact sheet).
1. Platform Defaults: Noise Suppression Is Now Built Into the Meeting Stack
The adoption question has largely been settled by defaults rather than purchasing decisions. Google’s Meet help center states that noise cancellation is on by default for Business and Enterprise accounts, while Education Fundamentals, Education Standard and Education Plus have it off by default (Google, Filter out background noise in Google Meet). Microsoft’s admin documentation lists the Teams VoiceIsolation meeting policy, plus the EnrollVoice and PassiveVoiceEnrollment voice-profile settings, as Enabled by default (Microsoft Learn, Manage voice isolation, updated April 2026). Voice isolation goes a step beyond noise suppression: it uses a voice profile stored on the user’s device to remove sounds and voices that do not match it.
The reach of those defaults is large. Microsoft reported more than 320 million monthly active Teams users and more than 10,000 paid Teams Premium customers in October 2023 (Microsoft, FY2024 Q1 earnings call). Most recent available data: Microsoft, 2023; its July 2026 call gave no new Teams user count. Krisp, the best-known standalone vendor, says its AI has processed trillions of minutes of voice conversations (Krisp, About us).
| Metric | Value | Source |
|---|---|---|
| Google Meet noise cancellation default | On for Business and Enterprise; off for Education editions | Google, Meet Help Center 2026 |
| Meet device-based noise cancellation on Android | Pixel 4 and later, Samsung S10 and later, other models | Google, Meet Help Center 2026 |
| Teams VoiceIsolation policy default | Enabled (EnrollVoice and PassiveVoiceEnrollment also Enabled) | Microsoft Learn, April 2026 |
| Teams monthly active users | More than 320 million | Microsoft, FY2024 Q1 earnings call (2023) |
| Teams Premium paid customers | More than 10,000 | Microsoft, FY2024 Q1 earnings call (2023) |
| Zoom background noise suppression levels | Auto (default), Low, Medium, High | Zoom Support, professional audio settings 2026 |
| Apple Voice Isolation mic mode | FaceTime iOS 15+, phone calls iOS 16.4+, Automatic mode iOS 18+ | Apple Support 2026 |
| Krisp voice minutes processed | Trillions | Krisp, About us 2026 |
Context note: Zoom’s High setting is described as its most aggressive, able to eliminate background speech from others, and may increase CPU use; its original sound mode for musicians disables post-processing and raises the codec to 48 kHz, 96 kbps mono or 192 kbps stereo (Zoom Support, professional audio settings). On the creator side, NVIDIA Broadcast’s Noise Removal and Room Echo Removal require a GeForce RTX 2060, Quadro RTX 3000, TITAN RTX or higher GPU (NVIDIA, Broadcast app). Apple’s mic mode requirements come from Apple Support.
2. The Cost of Bad Audio: Meeting Problems and Workplace Noise
Noise suppression sells because poor audio is the most common failure in hybrid meetings. In Owl Labs’ July 2025 survey of 2,000 US full-time workers, 76% cited audio echo or distortion and 75% cited not being able to hear everyone as common hybrid meeting issues, close behind lost time due to technical difficulties at 77% (Owl Labs, State of Hybrid Work 2025). Workers lose about 6 minutes on average getting hybrid meetings started, and 27% lose 10 or more minutes per meeting to tech setup.
Microsoft’s own research ties audio and video quality to whether meetings work for everyone. Across 3,290 email survey responses and 17,287 in-client ratings, quality, reliability and having an agenda could each improve the odds of a meeting being rated inclusive by 70% (Cutler et al., Meeting Effectiveness and Inclusiveness in Remote Collaboration, arXiv 2021). Most recent available data: Microsoft Research, 2021. The study measured audio and video quality together, so it cannot isolate the effect of background noise alone. Office acoustics, the other half of the problem, are covered in our open-plan office distraction and acoustics statistics.
| Metric | Value | Source |
|---|---|---|
| Workers citing audio echo or distortion as a common hybrid meeting issue | 76% | Owl Labs, State of Hybrid Work 2025 |
| Not being able to hear everyone | 75% | Owl Labs, State of Hybrid Work 2025 |
| Lost time due to technical difficulties | 77% | Owl Labs, State of Hybrid Work 2025 |
| Average time lost starting a hybrid meeting | About 6 minutes (27% lose 10+ minutes) | Owl Labs, State of Hybrid Work 2025 |
| Meeting survey sample (email / in-client) | 3,290 responses / 17,287 ratings (about 15% response rate in-client) | Cutler et al., Microsoft Research 2021 |
| Effect of AV quality, reliability or agenda on odds of inclusiveness | Each +70% | Cutler et al., Microsoft Research 2021 |
| US workers exposed to hazardous noise each year | About 27 million | NIOSH, CDC |
| US working population with hearing difficulty / tinnitus | About 1 in 9 / about 1 in 13 | NIOSH, CDC |
Context note: hazardous noise in NIOSH’s definition starts at 85 dBA, its recommended exposure limit averaged over an eight-hour day (NIOSH, About Noise and Hearing Loss), which is far louder than a typical office. A 2021 EEG study found brain activity partially confirmed a lower mental load with active noise-cancelling headphones than in noise, though the subjective difference was not significant (Pieper et al., Frontiers in Neuroscience 2021). Fatigue effects of video meetings are tracked in our video meeting fatigue statistics.
3. The Research Benchmark: Microsoft’s Deep Noise Suppression Challenge
Most of today’s real-time noise suppressors were shaped by one open benchmark. The first Deep Noise Suppression Challenge at Interspeech 2020 released 441 hours of clean speech from 2,150 speakers plus 60,000 noise clips across 150 audio classes, and drew 28 submissions from 19 teams (Reddy et al., Interspeech 2020 DNS Challenge). Its organizers stressed that the synthetic test set was not representative of real recordings, which is why every later edition leaned on blind real-world clips and crowdsourced ITU-T P.808 ratings.
The bar kept rising. By the fifth edition at ICASSP 2023, one track used 760.53 hours of clean speech (562.72 hours read speech, 185.41 hours Mandarin), 181 hours of noise in 152 classes, and models had to run at a real-time factor of 0.5 or less on a quad-core Intel Core i5 with combined latency of 20 ms or less (Dubey et al., ICASSP 2023 DNS Challenge; Microsoft Research challenge page). The task also shifted from removing noise to removing interfering talkers, and the paper reports degraded signal quality for most models, a sign that voice isolation is harder than noise removal.
| Metric | Value | Source |
|---|---|---|
| Clean speech released, 2020 | 441 hours from 2,150 speakers | Reddy et al., Interspeech 2020 |
| Participation, 2020 | 28 submissions from 19 teams (9 in both tracks) | Reddy et al., Interspeech 2020 |
| Crowdsourced ratings per clip, 2020 | 10 (20 for top entries) | Reddy et al., Interspeech 2020 |
| Clean speech, ICASSP 2023 Track 2 | 760.53 hours | Dubey et al., ICASSP 2023 |
| Noise data, 2023 | 181 hours, 152 audio classes, 60,000 clips | Dubey et al., ICASSP 2023 |
| Participation, 2023 | 11 submissions per track, 10 teams in both | Dubey et al., ICASSP 2023 |
| Best challenge-score gain over noisy blind set, 2023 | 0.145 (headset), 0.141 (speakerphone), same team won both | Dubey et al., ICASSP 2023 |
| Compute limits, 2023 | Real-time factor 0.5 or less; combined latency 20 ms or less | Microsoft Research, ICASSP 2023 DNS Challenge |
Context note: the public dataset is large: the challenge repository lists 827 GB of clean speech, 58 GB of noise and about 1 TB unpacked (Microsoft, DNS-Challenge repository). The 2022 edition added word accuracy as a metric and concluded that the research community still has “a long way to go” in challenging real-world noise (Dubey et al., ICASSP 2022 DNS Challenge). Most recent available challenge results: ICASSP 2023.
4. Voice Isolation and Speech Recognition Accuracy
Noise suppression now feeds machines as much as people, because meeting transcripts and voice agents depend on clean input. In Krisp’s 2026 benchmark, voice isolation cut word error rate in work environments from 31.84% to 6.30% across 11 speech-to-text configurations from 9 engines (Krisp, Voice Isolation Benchmark 2026). The gains were largest when a second voice was present; on clean phone calls, isolation slightly raised errors, from 3.48% to 3.91%.
That result matches the academic record. A 2026 paper notes that monaural speech enhancement “has not been established as an effective frontend” for recognition compared with models trained on noisy speech, then reports a 28.4% relative WER reduction on CHiME-2 by decoupling the two stages (Yang, Pandey and Wang, Computer Speech and Language 2026). The practical takeaway: isolation is a fix for crosstalk and background voices, not a free upgrade for every call.
| Metric | Value | Source |
|---|---|---|
| Overall WER change with voice isolation | -73% | Krisp, Voice Isolation Benchmark 2026 |
| Work environments (raw vs isolated) | 31.84% to 6.30% WER | Krisp, Voice Isolation Benchmark 2026 |
| Clean phone calls (raw vs isolated) | 3.48% to 3.91% WER (slight regression) | Krisp, Voice Isolation Benchmark 2026 |
| AssemblyAI Universal-3 Pro batch | 30.32% to 3.75% (-88%) | Krisp, Voice Isolation Benchmark 2026 |
| ElevenLabs Scribe v2 / Soniox STT v5 | -85% / -82% | Krisp, Voice Isolation Benchmark 2026 |
| Benchmark dataset | 265 conversations: 96 office, 106 call center, 63 mobile or car | Krisp, benchmark dataset post 2026 |
| Decoupled enhancement plus ASR on CHiME-2 | 28.4% relative WER reduction (5.6% WER) | Yang, Pandey and Wang, Computer Speech and Language 2026 |
| Same method, single-channel CHiME-4 | 3.3% WER simulated, 4.4% real | Yang, Pandey and Wang, Computer Speech and Language 2026 |
Context note: the Krisp benchmark is vendor-run and tests Krisp’s own models, although the dataset is published on Hugging Face. It covers 4 voice isolation models, 24 office devices, 18 call-center headset models and 18 phone models (Krisp, how the benchmark was recorded). Independent replication has not been published yet.
5. Contact Centers: Where Noise Removal Meets Voice AI
Call centers combine the noisiest rooms with the most voice-dependent workflows. In Krisp’s benchmark, call-center audio went from 23.83% to 6.86% word error rate after voice isolation, the setting where speech analytics, agent assist and translation all inherit the quality of the agent’s microphone feed (Krisp, Voice Isolation Benchmark 2026). The investment intent is clear: 60% of CX leaders plan to adopt AI voice translation within 6 to 12 months, up from 36% in 2025 (Krisp, The 2026 State of Voice in CX).
The agents’ own ears matter too. A Polish study of 78 call center operators measured average personal daily noise exposure of 74.7 dBA, with 92.3% having normal hearing in both ears (Pawlaczyk-Luszczynska et al., Noise and Health 2018). A larger 2022 study of 213 headset users, including 177 call center operators, found exposures ranging from 57 to 96 dB, though only 1.4% of operators exceeded 85 dB (Pawlaczyk-Luszczynska et al., IJOMEH 2022). Workload and attrition data for these teams are in our call center agent burnout and turnover statistics.
| Metric | Value | Source |
|---|---|---|
| Call-center WER, raw vs voice isolation | 23.83% to 6.86% | Krisp, Voice Isolation Benchmark 2026 |
| CX leaders planning AI voice translation within 6-12 months | 60% (36% in 2025) | Krisp, 2026 State of Voice in CX (815 leaders, 12 countries) |
| Accent conversion, current or planned adoption | 61% | Krisp, 2026 State of Voice in CX |
| Speech analytics adoption | 80% | Krisp, 2026 State of Voice in CX |
| Top challenge: hiring agents in the right regions and languages | 40% (language support costs 37%; tech integration 14%) | Krisp, 2026 State of Voice in CX |
| Average daily noise exposure, 78 call center operators | 74.7 dBA (range 68-79) | Pawlaczyk-Luszczynska et al., Noise and Health 2018 |
| Headset users exceeding 85 dB, 177 operators | 1.4% | Pawlaczyk-Luszczynska et al., IJOMEH 2022 |
| Headset users with bilateral normal hearing (250-8000 Hz) | 42% of 213 | Pawlaczyk-Luszczynska et al., IJOMEH 2022 |
Context note: the State of Voice in CX survey is published by a vendor that sells voice translation and accent conversion, so its adoption figures should be read as directional. The 2018 study is the most recent available data on call center operator exposure in that cohort; it also found high-frequency notches in 15.4% of audiograms.
6. Beyond Meetings: Deep Learning Noise Reduction in Hearing Care
The same neural denoising ideas now ship in hearing aids, where the stakes are higher than a cleaner transcript. In a 2025 study of 11 bimodal cochlear implant users aged 71-89, a DNN hearing aid program raised sentence recognition in noise to 79%, compared with 60% using a conventional setting (p < 0.001), and offered a 40% bimodal benefit over the implant alone versus 21% (Kolberg et al., Journal of Clinical Medicine 2025). Field data point the same way: in 40 experienced users who logged 3,614 momentary assessments and 6,812 hours of sound, the DNN-based hearing aid kept satisfaction consistent regardless of background noise level (Christensen et al., American Journal of Audiology 2024).
The addressable population is large. The WHO estimates that more than 430 million people need rehabilitation for disabling hearing loss and that nearly 2.5 billion will have some degree of hearing loss by 2050 (WHO, Deafness and hearing loss fact sheet). A 2023 study reported a single-microphone deep learning denoiser that restored speech intelligibility for hearing aid users to the level of normal-hearing controls and ran in real time on a laptop (Diehl et al., Scientific Reports 2023).
| Metric | Value | Source |
|---|---|---|
| Sentence recognition in noise, DNN vs conventional program | 79% vs 60% (p < 0.001), 11 users aged 71-89 | Kolberg et al., J Clin Med 2025 |
| Bimodal benefit over implant alone in babble | 40% (DNN) vs 21% | Kolberg et al., J Clin Med 2025 |
| Real-world DNN hearing aid trial | 40 users, 3,614 assessments, 6,812 hours of audio | Christensen et al., Am J Audiol 2024 |
| DNN combined with beamforming in older adults | Outperformed other programs on all metrics (20 participants) | Folkeard et al., Trends in Hearing 2026 |
| Noise reduction in severe-to-profound hearing loss | Significantly reduced listening effort (25 participants) | Dong et al., Frontiers in Neuroscience 2024 |
| Objective DNN evaluation | 1,152 manikin recordings at -5, 0 and +5 dB SNR | Ashkanichenarlogh et al., Trends in Hearing 2025 |
| People needing rehabilitation for disabling hearing loss | More than 430 million (over 5% of the world) | WHO fact sheet |
| Annual global cost of unaddressed hearing loss | Almost US$1 trillion | WHO fact sheet |
Context note: these trials are small, from 11 to 40 participants, and several were run with a single manufacturer’s devices. The Folkeard, Dong and Ashkanichenarlogh results are summarized from their abstracts (Folkeard et al. 2026; Dong et al. 2024; Ashkanichenarlogh et al. 2025), and benefits in the 2025 objective study were attenuated under moderate reverberation.
Summary: AI Noise Cancellation Software by the Numbers
| Metric | Value | Source |
|---|---|---|
| Overall WER change with voice isolation | -73% across 265 conversations | Krisp, Voice Isolation Benchmark 2026 |
| Work-environment WER, raw vs isolated | 31.84% to 6.30% | Krisp, Voice Isolation Benchmark 2026 |
| Clean phone call WER, raw vs isolated | 3.48% to 3.91% | Krisp, Voice Isolation Benchmark 2026 |
| Call-center WER, raw vs isolated | 23.83% to 6.86% | Krisp, Voice Isolation Benchmark 2026 |
| Google Meet noise cancellation | On by default for Business and Enterprise | Google, Meet Help Center 2026 |
| Teams VoiceIsolation policy | Enabled by default | Microsoft Learn, April 2026 |
| Teams monthly active users | More than 320 million (2023) | Microsoft, FY2024 Q1 earnings call |
| Zoom noise suppression levels | 4 (Auto default, Low, Medium, High) | Zoom Support 2026 |
| Audio echo or distortion as common hybrid meeting issue | 76% | Owl Labs, State of Hybrid Work 2025 |
| Not being able to hear everyone | 75% | Owl Labs, State of Hybrid Work 2025 |
| US workers exposed to hazardous noise | About 27 million a year | NIOSH, CDC |
| DNS Challenge 2020 clean speech / submissions | 441 hours / 28 from 19 teams | Reddy et al., Interspeech 2020 |
| DNS Challenge 2023 clean speech, Track 2 | 760.53 hours | Dubey et al., ICASSP 2023 |
| DNS Challenge 2023 latency limit | 20 ms or less | Microsoft Research 2023 |
| Relative WER reduction, decoupled enhancement on CHiME-2 | 28.4% | Yang, Pandey and Wang 2026 |
| CX leaders planning AI voice translation | 60% (36% in 2025) | Krisp, 2026 State of Voice in CX |
| Call center operators’ average daily noise exposure | 74.7 dBA | Pawlaczyk-Luszczynska et al. 2018 |
| DNN hearing aid sentence recognition in noise | 79% vs 60% | Kolberg et al. 2025 |
| People needing hearing rehabilitation | More than 430 million | WHO fact sheet |
Methodology and Sources
Every figure above was read during research for this article in a company document, help center page, earnings transcript, government page, challenge paper or peer-reviewed abstract. Vendor benchmarks and surveys are labeled as such.
- Krisp: About us, Voice Isolation Benchmark, September 2026, benchmark dataset recording post, September 2026, The 2026 State of Voice in CX, August 2026
- Microsoft: FY2024 Q1 earnings call, October 2023, Microsoft Learn, Manage voice isolation (April 2026)
- Microsoft Research DNS Challenge: Interspeech 2020 paper, ICASSP 2022 paper, ICASSP 2023 paper, ICASSP 2023 challenge page, DNS-Challenge repository
- Microsoft Research meeting study: Cutler et al., Meeting Effectiveness and Inclusiveness in Remote Collaboration, 2021
- Google: Filter out background noise in Google Meet
- Zoom: Professional audio settings support article
- Apple: Mic modes support article
- NVIDIA: Broadcast app
- Owl Labs: State of Hybrid Work 2025
- NIOSH (CDC): About Noise and Hearing Loss
- World Health Organization: Deafness and hearing loss fact sheet
- Peer-reviewed research via Europe PMC: Yang, Pandey and Wang 2026, Pieper et al. 2021, Pawlaczyk-Luszczynska et al. 2018, Pawlaczyk-Luszczynska et al. 2022, Kolberg et al. 2025, Christensen et al. 2024, Diehl et al. 2023
- Data watch: No market-size figure for AI noise cancellation software is included, because we could not read two independent research firms’ estimates in primary form and would not cite one firm alone. Krisp is private and publishes no user count; its “trillions of minutes” claim is a company statement, and both its benchmark and its CX survey are vendor-run. The Teams user count (320 million) dates from October 2023 and Microsoft’s July 2026 call gave no update. The Microsoft Teams blog post announcing AI noise suppression in 2020 and the Webex noise removal help article could not be loaded, so Cisco Webex is not covered. The 2023 DNS Challenge paper gives score gains but its full results table could not be extracted. The Cutler et al. meeting study (2021), the 2018 call center exposure study and the 2023 DNS Challenge are the most recent available data of their kind. Hearing aid trials are small (11 to 40 participants) and partly manufacturer-linked. The Owl Labs meeting-issue percentages are self-reported by US full-time workers surveyed in July 2025.
Last updated: October 3, 2026. We update this roundup quarterly, and the next refresh is expected when Owl Labs publishes its State of Hybrid Work 2026 report and Microsoft announces results of its next Deep Noise Suppression or speech quality challenge.