Whisper large-v2, a widely used speech recognition model, made 36.4% word errors on meeting audio picked up by a single distant microphone, against 16.9% on the same meetings recorded through headset microphones (OpenAI, Whisper paper 2022). The gap starts at capture: a laptop’s built-in mic at about 60 cm recorded speech 12.82 dB lower than a headset mic at 15 to 20 cm in a 30-person study (Busquet et al., Behavior Research Methods 2023). And it does not stay technical: 600 listeners rated a job candidate heard through a tinny, poor-microphone sound 68.50 out of 100 for hireability, versus 76.04 for the clear recording (Walter-Terrill et al., PNAS 2025). We aggregated data from Amazon, Yale University, the REVERB and CHiME challenges, Ear and Hearing, a five-organization meeting study, ITU-T and Zoom, plus two vendor surveys labeled as such. For the hardware market itself, see our webcam market statistics; this page is about what the microphone does to the voice.
TL;DR
- Whisper large-v2 word error rate on AMI meetings: 16.9% with individual headset mics vs 36.4% with one distant mic (OpenAI, Whisper paper 2022).
- A laptop built-in mic at about 60 cm captured speech at 46.91 dB vs 59.73 dB for a headset mic at 15 to 20 cm (Busquet et al., Behavior Research Methods 2023).
- On Amazon’s DiPCo dinner-party recordings, close-talk WER was 42.10% vs 77.52% for far-field arrays (Van Segbroeck et al., DiPCo 2019).
- AliMeeting Mandarin meetings: 9.9% character errors on headset audio vs 49.0% on far-field array audio (Yu et al., M2MeT 2022).
- Hireability fell from 76.04 to 68.50 out of 100 when the same voice sounded like a poor microphone, with transcription accuracy unchanged (98.09% vs 98.07%) (Walter-Terrill et al., PNAS 2025).
- Low-quality audio scored a ViSQOL MOS of 2.13 vs 4.68 for medium quality, and lowered employability ratings (Schütze et al., Forum Acusticum 2026).
- Background noise slowed listener responses by a mean of 91 ms, a measure of listening effort (Picou et al., Ear and Hearing 2016).
- Adults with cochlear implants reached about 60% of typical listening efficiency in audio-only online conversations, rising to about 80% with captions (Perea Pérez et al., Ear and Hearing 2025).
- Only 30% of knowledge workers use a professional headset for work and 8% call it their main device (Jabra 2023 Knowledge Worker Study, vendor survey).
1. Mic Distance and Captured Speech Level
A webcam or laptop microphone is a far-field microphone: it sits roughly an arm’s length from the mouth or further, so it records the room along with the voice. A headset mic is a close-talk microphone, a few centimetres from the lips. The cleanest measured comparison comes from a 2023 study that recorded the same speakers on several devices. The laptop’s built-in mic at about 60 cm captured speech at 46.91 dB, while a headset mic at 15 to 20 cm captured 59.73 dB, a 12.82 dB difference (59.73 minus 46.91) (Busquet, Efthymiou and Hildebrand, Behavior Research Methods 2023).
Two details matter for buyers. First, even a studio microphone at the same 60 cm out-captured the laptop by 10.76 dB (57.67 minus 46.91), so the built-in mic loses ground through both distance and hardware. Second, the laptop and the studio mic measured average pitch almost identically, which means a far mic does not change who you sound like; it changes how much of you reaches the other side compared with the room. A separate test of group-work recordings shows the other half of the story: with six different microphones all placed 5 feet (about 1.5 m) away, model choice barely moved the error rate (Bradford et al., EDM 2022).
| Metric | Value | Source |
|---|---|---|
| Speech level, laptop built-in mic at about 60 cm vs headset mic at 15 to 20 cm | 46.91 dB vs 59.73 dB | Busquet et al., Behavior Research Methods 2023 |
| Speech level, studio microphone at about 60 cm vs lavalier (clip-on) mic at 15 to 20 cm | 57.67 dB vs 63.68 dB | Busquet et al., Behavior Research Methods 2023 |
| Average voice pitch measured by the laptop mic vs the studio microphone | 168.42 Hz vs 169.31 Hz (t = -.89, p = .99) | Busquet et al., Behavior Research Methods 2023 |
| Size of the device comparison study | 30 participants, 1,800 voice samples | Busquet et al., Behavior Research Methods 2023 |
| Word error rate, six microphones 5 feet from speakers, read-aloud script | 10.7% to 12.0% | Bradford et al., EDM 2022 |
| Word error rate, same six microphones, free group conversation | 52.7% to 61.3% | Bradford et al., EDM 2022 |
| Distance from each speaker to the far-field arrays in the DiPCo recordings | 1,120 mm to 3,950 mm | Van Segbroeck et al., DiPCo 2019 |
| Microphone-to-speaker distance in the AliMeeting meeting rooms | 0.3 to 5.0 m | Yu et al., M2MeT 2022 |
Context note: the decibel values in that study are amplitude measures from the recordings, not calibrated sound pressure levels, so the 12.82 dB gap compares devices within one setup rather than predicting what any webcam will deliver. Most recent available data for the EDM comparison: 2022.
2. Speech Recognition Errors: Close-Talk vs Far-Field
Speech recognition is the most measurable listener there is, and it says the same thing in every corpus: move the mic away and errors roughly double or worse. Whisper large-v2 went from 16.9% WER (word error rate, the share of words transcribed wrongly) on AMI meetings recorded with individual headset mics to 36.4% on the same meetings from a single distant mic, about 2.2 times higher (36.4 divided by 16.9) (Radford et al., OpenAI Whisper paper 2022). An older wav2vec 2.0 model shows the same direction, 37.0% vs 67.6%.
The pattern holds in homes and in other languages. Amazon’s DiPCo corpus recorded four people around a dining table with headsets and five far-field arrays at the same time: the best evaluation setup reached 42.10% close-talk WER and 77.52% far-field, about 1.8 times higher (77.52 divided by 42.10) (Van Segbroeck et al., DiPCo 2019). In Mandarin meetings the gap was steeper, 9.9% vs 49.0% CER (character error rate, the Mandarin equivalent of WER), about 5 times (49.0 divided by 9.9) (Yu et al., M2MeT 2022). Training on far-field audio narrowed it to 8.2% vs 33.3% but did not close it. Rooms with ceiling arrays have their own data, covered in our conference room beamforming microphone statistics.
| Metric | Value | Source |
|---|---|---|
| Whisper large-v2 WER, AMI meetings: individual headset mics vs single distant mic | 16.9% vs 36.4% | Radford et al., OpenAI Whisper paper 2022 |
| wav2vec 2.0 Large (no language model) WER, AMI: headset mics vs single distant mic | 37.0% vs 67.6% | Radford et al., OpenAI Whisper paper 2022 |
| REVERB real recordings WER, best eight-microphone system | 9.0% | Kinoshita et al., REVERB challenge summary 2016 |
| DiPCo evaluation WER, close-talk headset vs far-field array (best language model) | 42.10% vs 77.52% | Van Segbroeck et al., DiPCo 2019 |
| AliMeeting CER, model trained on headset audio: headset test vs far-field test | 9.9% vs 49.0% | Yu et al., M2MeT 2022 |
| AliMeeting CER, model adding far-field, external and simulated data: headset vs far-field | 8.2% vs 33.3% | Yu et al., M2MeT 2022 |
| REVERB real recordings WER: clean/headset reference vs far-field baseline | 6.9% vs 49.2% | Kinoshita et al., REVERB challenge summary 2016 |
| CHiME-6 official baseline WER on far-field arrays, development vs evaluation set | 51.8% vs 51.3% | Watanabe et al., CHiME-6 challenge 2020 |
Most recent available data: the Whisper paper (2022) is still the largest public side-by-side of headset vs distant-mic WER on the same meetings; newer models have not published an AMI headset vs distant split in the same format. The CHiME-6 baseline paper only allowed far-field arrays at test time, which is why it has no close-talk column.
3. Rooms, Noise and Listening Effort
Distance matters because of what fills the gap: reflections and noise. A typical living room takes 0.3 to 0.7 s for its response to decay by 60 dB (the reverberation time), according to a review of far-field recognition (Haeb-Umbach et al., Far-Field Automatic Speech Recognition 2020). A headset mic close to the mouth hears much less of that tail than a webcam mic on top of a monitor.
For human listeners the evidence splits in a useful way. In a controlled Ear and Hearing study, background noise slowed responses by a mean of 91 ms (the study’s measure of listening effort), while reverberation did not significantly change effort, even though each step up in reverberation cut word recognition by about 8 RAU (rationalized arcsine units, a transformed percent-correct score) (Picou, Gordon and Ricketts, Ear and Hearing 2016). In other words, room echo costs words; room noise costs energy. That matters because a far mic picks up more of both. For people with hearing loss, Australia’s National Acoustic Laboratories found that following videoconference conversations required a lot more effort than for people with normal hearing, and hearing aids did not restore face-to-face levels. The physiological side of call fatigue is covered in our video meeting fatigue biomarker statistics.
| Metric | Value | Source |
|---|---|---|
| Typical living-room reverberation time (decay to -60 dB) | 0.3 to 0.7 s | Haeb-Umbach et al., Far-Field ASR review 2020 |
| Word recognition loss per step up in reverberation, at a fixed signal-to-noise ratio | about 8 RAU | Picou et al., Ear and Hearing 2016 |
| Response-time slowdown in noise vs quiet (listening effort) | mean 91 ms | Picou et al., Ear and Hearing 2016 |
| Effect of reverberation on listening effort when recognition was comparable | not significant (p = 0.18) | Picou et al., Ear and Hearing 2016 |
| Reverberation times tested in that study | T30 under 100 ms, 475 ms and 834 ms | Picou et al., Ear and Hearing 2016 |
| REVERB simulated rooms: reverberation times, and near vs far mic distance | 0.3, 0.6 and 0.7 s; 50 cm vs 200 cm | Kinoshita et al., REVERB challenge summary 2016 |
| CHiME-5 far-field WER: baseline vs strong back-end vs multi-microphone separation and dereverberation | barely below 80% vs approximately 60% vs 43.2% | Haeb-Umbach et al., Far-Field ASR review 2020 |
| Listening efficiency of cochlear implant users in online conversations, audio only vs with captions (share of the typical-hearing mean) | about 60% vs about 80% | Perea Pérez et al., Ear and Hearing 2025 |
Outlier note: the null effect of reverberation on effort surprised the authors themselves, who called it inconsistent with current models of listening effort; it was measured in 18 young adults with normal hearing. Most recent available data: 2016. Noise is the half that software can attack, see our neural noise suppression statistics.
4. How Listeners Judge a Speaker Through a Poor Mic
The most consequential finding of the last two years is that bad mic sound changes judgments of the person, not just intelligibility. In a PNAS experiment with 600 listeners, a job candidate heard through a tinny, hollow ‘poor microphone’ sound received a hireability score of 68.50 vs 76.04 for the clear recording (d = 0.35, P < 0.001), while transcription accuracy was identical at 98.07% vs 98.09% (Walter-Terrill, Ongchoco and Scholl, PNAS 2025). Listeners understood every word; they simply thought less of the speaker.
A 2026 follow-up with 299 participants and avatar job interviews found that low-bitrate, band-limited audio lowered employability ratings (p < 0.001), while the step from medium to high quality made no difference (p = 0.813) (Schütze, Biberger and Reber, Forum Acusticum 2026). That suggests a threshold: listeners punish clearly degraded audio, not merely ordinary audio. Earlier work found the same bias in simulated Skype interviews, where it persisted even after raters were told to ignore call quality (Fiechter et al., Cognitive Research 2018).
| Metric | Value | Source |
|---|---|---|
| Hireability score (0 to 100), clear recording vs tinny ‘poor microphone’ version | 76.04 vs 68.50 (d = 0.35) | Walter-Terrill et al., PNAS 2025 |
| Transcription accuracy, clear vs distorted recording | 98.09% vs 98.07% (P = 0.915) | Walter-Terrill et al., PNAS 2025 |
| Listeners in Experiment 1 | 600 (300 per condition) | Walter-Terrill et al., PNAS 2025 |
| Employability rating gap, high vs low audio quality (0 to 100 scale) | 6.4 points; low vs high p < 0.001; medium vs high p = 0.813 | Schütze et al., Forum Acusticum 2026 |
| Objective quality (ViSQOL MOS) of low-quality vs medium-quality audio | 2.13 vs 4.68 | Schütze et al., Forum Acusticum 2026 |
| Hirability score (1 to 10), smooth vs degraded audio and video, both experiments combined | 6.91 vs 6.33 (Cohen’s d = 0.44) | Fiechter et al., Cognitive Research 2018 |
| Listeners who rated identical talks or radio interviews heard in good vs poor audio | 97 (Experiment 1) and 99 (Experiment 2) | Newman and Schwarz, Science Communication 2018 |
Context note: in the Newman and Schwarz study (DOI 10.1177/1075547018759345), poor audio lowered ratings of the talk, of the speaker’s intelligence and of the importance of the research, as summarized by USC. MOS (mean opinion score) is the standard listening-quality scale, where higher is better; ViSQOL is a software model that predicts it. Most recent available data for the Fiechter and Newman studies: 2018.
5. Certification Rules: What Teams, Zoom and ITU Actually Test
Certification programs reveal where vendors put the audio bar, and webcams sit mostly outside it. The Teams certified-device page says audio devices such as headsets and speakerphones must offer wide-band audio with no echo, distortion or excessive glitches, while the criteria it lists for video devices are all image metrics (Microsoft Learn, Teams certified USB devices, November 2026). Of the 22 USB devices on that list qualified in 2026, 2 are webcams and 20 are headsets or headset systems (our count of the November 2026 table).
The headset bar also has a premium tier. Test-lab documentation for the Teams Audio Test Specification describes an optional open-office module that scores speech and noise MOS under office noise plus “distractor attenuation”, meaning how much of a neighbour’s voice the headset rejects (HEAD acoustics datasheet, Teams Audio Test Specification 60026). Zoom unified the objective testing standards for its audio categories in 2022 (Zoom, hardware certification update 2022), and the Teams lab setup builds on ITU-T methods such as P.502 and the P.51 artificial mouth (a loudspeaker that imitates a human talker). The wider reference for scoring captured speech is ITU-T P.863 (POLQA), the standard model that predicts listening quality on the MOS scale.
| Metric | Value | Source |
|---|---|---|
| Audio requirement for Teams-certified headsets and speakerphones | wide-band audio with no echo, distortion or excessive glitches; echo cancellation supported | Microsoft Learn, Teams certified USB devices 2026 |
| Criteria listed for Teams-certified video devices such as webcams | noise, color, image detail, jitter, latency, frame rate, field of view (image metrics) | Microsoft Learn, Teams certified USB devices 2026 |
| USB devices qualified in 2026 (up to the November 2026 table): webcams vs headsets or headset systems | 2 vs 20 (of 22) | Microsoft Learn, our count of the November 2026 table |
| Teams Audio Test Specification version used by test labs, and device categories | V5.0 including Addendum A; 4 categories (headsets, personal desk phones, personal speakerphones, shared speakerphones) | HEAD acoustics test-lab datasheet 2026 |
| Extra metrics in the optional open-office headset test | 3QUEST speech and noise MOS; distractor attenuation (average and minimum) | HEAD acoustics test-lab datasheet 2026 |
| Zoom audio and related categories with refreshed, unified objective test standards | 7 | Zoom hardware certification update, March 16, 2022 |
| How often Zoom-certified devices must be recertified | at least every two years | Zoom hardware certification update, March 16, 2022 |
| ITU-T P.863 (POLQA) edition in force, plus its amendment | P.863 (03/18); Amendment 1 (04/20) on predicting acoustically recorded narrowband speech | ITU-T Recommendation P.863 |
Context note: “wide-band” here is a telephony term for audio that keeps a wider range of frequencies than traditional narrowband phone audio; it is not a measure of loudness. Most recent available data for Zoom’s criteria: 2022, as Zoom has not published newer thresholds.
6. How People Equip Themselves and What It Changes in Real Meetings
Survey data on equipment is thin and mostly vendor-run, but it points one way: most people still talk through whatever is built in. Only 30% of knowledge workers use a professional headset for work, and 8% call it their main device, although 91% use an online meeting platform (Jabra 2023 Knowledge Worker Study, vendor survey). A 2022 Logitech survey of more than 3,000 remote workers in the US, Germany and China found 85% struggle with sound problems (Logitech, vendor survey).
Independent data adds the outcome side. A study of about 7,330 usable post-meeting survey responses from five large organizations found that using a headset raised the odds of a ‘Participation’ rating (OR = 1.16), and calls with audio quality issues had far lower odds of being rated effective (OR = 0.14) (Hosseinkashi et al., Microsoft Research, meeting effectiveness study 2024). An odds ratio (OR) below 1 means lower odds; 0.14 is a large effect. Read with the level gap in section 1, the case for close-talk capture is about being heard as fully present, not just being heard.
| Metric | Value | Source |
|---|---|---|
| Knowledge workers using a professional headset for work | 30% | Jabra 2023 Knowledge Worker Study (vendor survey, 7,000+ employees) |
| Knowledge workers who say a professional headset is their main device | 8% | Jabra 2023 Knowledge Worker Study (vendor survey) |
| Remote workers who struggle with sound problems on calls | 85% | Logitech end-user survey, May 2022 (vendor survey, 3,000+ workers) |
| Online-conversation study, typical-hearing group: listened on loudspeakers vs headphones | 72% vs 28% | Perea Pérez et al., Ear and Hearing 2025 |
| Same study, cochlear implant group: loudspeakers vs headphones vs direct streaming | 68% vs 12% vs 20% | Perea Pérez et al., Ear and Hearing 2025 |
| Adjusted odds ratio for a ‘Participation’ rating when the user wears a headset | 1.16 | Hosseinkashi et al., Microsoft Research 2024 |
| Participation rate, audio-only calls with a headset vs without | about 5% higher | Hosseinkashi et al., Microsoft Research 2024 |
| Adjusted odds ratio for an ‘Effective’ meeting rating when the call had quality issues (echo, noise, distortion) | 0.14 | Hosseinkashi et al., Microsoft Research 2024 |
Outlier note: the meeting study’s text says the odds of Participation rise “by 20%” with a headset, while its own table gives an odds ratio of 1.16; we report the table value. The authors also warn that headset detection from telemetry is unreliable because many devices report it incorrectly. Most recent available data for the Logitech survey: May 2022.
Summary: Webcam Integrated Microphone vs Headset Speech Quality by the Numbers
| Metric | Value | Source |
|---|---|---|
| Whisper large-v2 WER, AMI headset mics vs single distant mic | 16.9% vs 36.4% | Radford et al., OpenAI Whisper paper 2022 |
| DiPCo WER, close-talk vs far-field array | 42.10% vs 77.52% | Van Segbroeck et al., DiPCo 2019 |
| AliMeeting CER, headset vs far-field | 9.9% vs 49.0% | Yu et al., M2MeT 2022 |
| REVERB real recordings: clean/headset WER vs far-field baseline vs best 8-mic system | 6.9% vs 49.2% vs 9.0% | Kinoshita et al., REVERB challenge summary 2016 |
| Speech level, laptop mic at about 60 cm vs headset mic at 15 to 20 cm | 46.91 dB vs 59.73 dB | Busquet et al., Behavior Research Methods 2023 |
| WER range of six mics at 5 feet, script vs free conversation | 10.7% to 12.0% vs 52.7% to 61.3% | Bradford et al., EDM 2022 |
| Typical living-room reverberation time | 0.3 to 0.7 s | Haeb-Umbach et al., Far-Field ASR review 2020 |
| Response slowdown from background noise | mean 91 ms | Picou et al., Ear and Hearing 2016 |
| Cochlear implant listening efficiency online, audio only vs captions | about 60% vs about 80% of typical | Perea Pérez et al., Ear and Hearing 2025 |
| Hireability, clear vs poor-microphone sound (0 to 100) | 76.04 vs 68.50 | Walter-Terrill et al., PNAS 2025 |
| ViSQOL MOS, low vs medium audio quality | 2.13 vs 4.68 | Schütze et al., Forum Acusticum 2026 |
| Hirability (1 to 10), smooth vs degraded audio and video | 6.91 vs 6.33 | Fiechter et al., Cognitive Research 2018 |
| Professional headset use for work | 30% | Jabra 2023 Knowledge Worker Study (vendor survey) |
| Professional headset named as main device | 8% | Jabra 2023 Knowledge Worker Study (vendor survey) |
| Online-conversation listeners with typical hearing on loudspeakers vs headphones | 72% vs 28% | Perea Pérez et al., Ear and Hearing 2025 |
| Microphone-to-speaker distance in AliMeeting rooms | 0.3 to 5.0 m | Yu et al., M2MeT 2022 |
| Remote workers struggling with sound problems | 85% | Logitech end-user survey 2022 (vendor survey) |
| Zoom recertification interval | at least every two years | Zoom hardware certification update 2022 |
Methodology and Sources
- Busquet, Efthymiou and Hildebrand, Voice analytics in the wild: Validity and predictive accuracy of common audio-recording devices, Behavior Research Methods 2023
- Bradford et al., A deep dive into microphone hardware for recording collaborative group work, EDM 2022
- Radford et al. (OpenAI), Robust Speech Recognition via Large-Scale Weak Supervision (Whisper), 2022
- Van Segbroeck et al. (Amazon), DiPCo: Dinner Party Corpus, 2019
- Yu et al., M2MeT: The ICASSP 2022 Multi-channel Multi-party Meeting Transcription Challenge, 2022
- Kinoshita et al., A summary of the REVERB challenge, EURASIP Journal on Advances in Signal Processing 2016
- Watanabe et al., CHiME-6 Challenge: Tackling Multispeaker Speech Recognition for Unsegmented Recordings, 2020
- Haeb-Umbach et al., Far-Field Automatic Speech Recognition, 2020
- Picou, Gordon and Ricketts, The effects of noise and reverberation on listening effort in adults with normal hearing, Ear and Hearing 2016
- Perea Pérez et al., Preferences and Listening Efficiency of Adults With Cochlear Implants During Online Communication, Ear and Hearing 2025
- National Acoustic Laboratories, structured assessment of communication via videoconferencing, 2022
- Walter-Terrill, Ongchoco and Scholl, Superficial auditory (dis)fluency biases higher-level social judgment, PNAS 2025
- Schütze, Biberger and Reber, Subjective and objective employability for degraded speech in video-mediated job interviews, Forum Acusticum 2026 (preliminary paper)
- Fiechter et al., Audiovisual quality impacts assessments of job candidates in video interviews, Cognitive Research: Principles and Implications 2018
- Newman and Schwarz, Good Sound, Good Research, Science Communication 2018, as reported by USC
- Microsoft Learn, Teams certified USB devices table, November 2026
- Hosseinkashi et al. (Microsoft Research), Meeting effectiveness and inclusiveness: large-scale measurement in real-world remote meetings, 2024 revision
- HEAD acoustics, ACQUA Standards datasheet for the Teams Audio Test Specification 60026, 2026
- Zoom, New updates to the Zoom hardware certification program, March 16, 2022
- ITU-T Recommendation P.863, Perceptual objective listening quality prediction
- Jabra, LSE engagement study and 2023 Knowledge Worker Study (vendor research)
- Logitech, Your Remote Employees are Underequipped (vendor survey), with fieldwork by 2CV Research in May 2022 per the full Logitech study PDF
- Data watch: no public study has yet measured a current webcam’s built-in mic against a headset on the same speakers with POLQA or ASR, so this roundup combines laptop-mic, distant-mic and array data as the closest far-field evidence. Several key datasets are older than three years (REVERB 2016, the 2016 noise and reverberation study, Fiechter and Newman 2018, DiPCo 2019, Whisper and M2MeT 2022, Zoom criteria 2022) and are flagged in their sections. Conflicts and inconsistencies: the meeting study’s text gives a 20% rise in Participation odds for headset users while its table gives an odds ratio of 1.16 (we use the table); the DiPCo paper gives session lengths of 15 to 45 minutes in its abstract and 15 to 47 minutes in the body; the M2MeT paper prints two different overlap ratios for its evaluation set (we use neither). The 6.4-point employability gap comes from a preliminary Forum Acusticum paper that is not yet in a journal. Jabra and Logitech figures are vendor surveys and are labeled as such. The 12.82 dB, 10.76 dB and ratio figures are our own subtraction or division of published values, with the math shown in the text.
Last updated: November 7, 2026. We update this roundup quarterly, and the next refresh is expected when the final proceedings of Forum Acusticum 2026 (the 12th Convention of the European Acoustics Association) are published.