The official DIHARD III baseline detector missed 4.97% of speech on the evaluation core set but wrongly flagged non-speech as speech at a 15.07% rate, three times its own miss rate. That gap is the whole story of voice activity detection (VAD): every detector trades misses (real speech it drops) against false alarms (silence, music or noise it lets through), and benchmarks decide which error matters more. In the Fearless Steps Phase-3 challenge on NASA Apollo audio, the best system cut the baseline’s 15.16% detection cost to 1.47%, while at 0 dB signal-to-noise ratio WebRTC VAD averaged an F1 score of only 0.581. Upstream of everything, noise is the main enemy, which is why VAD and neural noise suppression latency are usually tuned together. We aggregated data from NIST’s OpenSAT evaluations, the DIHARD III and Fearless Steps challenges, Google’s AVA-Speech and 2026 window-size studies, 3GPP and ITU-T codec specifications, Cisco and Avaya telephony guidance, and more than a dozen peer-reviewed and preprint papers.
TL;DR
- NIST scored speech activity detection as 0.75 x miss rate + 0.25 x false alarm rate, making a miss three times as costly (NIST IR 8242, OpenSAT 2017 pilot).
- The DIHARD III baseline had 4.97% miss and 15.07% false alarm on the evaluation core set (Third DIHARD Diarization Challenge paper, arXiv 2020).
- The best Fearless Steps Phase-3 system reached a 1.47% detection cost on evaluation data against a 15.16% baseline (Joglekar et al., Interspeech 2021).
- On DIHARD 3 audio, pyannote 2.0 had 3.9% false alarm and 3.3% miss, against 17.0% and 4.0% for the Silero VAD available in 2021 (Bredin and Laurent, arXiv 2021).
- On the ESC-50 environmental noise set, WebRTC VAD scored 0 accuracy, Silero v5 0.61 and Silero v6 0.87 (Silero VAD Quality Metrics, project claim).
- At 50 ms windows, Silero’s Matthews correlation coefficient was 0.72 against 0.41 for WebRTC and 0.11 for a simple loudness (RMS) detector (McKinnon et al., Google, arXiv 2026).
- At 0 dB SNR, WebRTC VAD averaged an AUC of 0.691 against 0.956 for the 22.7K-parameter MagicNet (Jia et al., arXiv 2024).
- WebRTC VAD took 0.06 ms per decision, a large ResNet VAD 265 ms (Chaudhuri et al., AVA-Speech, Interspeech 2018).
- Cisco plans for VAD to save up to 35 percent of voice bandwidth, but only averaged over more than 24 calls (Cisco, Per Call Bandwidth Consumption).
- Whisper’s word error rate on noise-padded speech was 68.3% behind WebRTC VAD and 8.0% behind Silero VAD (Barański et al., arXiv 2025).
1. How VAD Errors Are Scored: Misses vs False Alarms
Before comparing any two VAD numbers, check the scoring rules. A “false alarm” is non-speech labeled as speech; a “miss” is speech labeled as non-speech. Evaluators then decide how much each error costs and how much slack to give around segment edges. That slack is called a “collar”: a short window around each true speech boundary where errors are not counted. NIST weighted misses at 0.75 and false alarms at 0.25, reasoning that “missed detections are considered more critical than false alarms” (NIST IR 8242, OpenSAT 2017 pilot).
The trend since then is toward stricter scoring. Collars shrank from 0.5 s to 0.25 s in Fearless Steps and disappeared entirely in DIHARD III, so a 2026 number scored with no collar is not comparable with a 2017 number scored with one. The NIST series itself is finished: the OpenSAT program page lists its status as completed, with no evaluation after 2020.
| Metric | Value | Source |
|---|---|---|
| Detection cost function weights used to score speech activity detection | DCF = 0.75 x P_Miss + 0.25 x P_FA | NIST IR 8242, OpenSAT 2017 pilot |
| Non-scored window (collar) around each reference speech segment, OpenSAT 2017 pilot | 0.5 s | NIST IR 8242, OpenSAT 2017 pilot |
| Teams that submitted results to the OpenSAT 2017 pilot vs sites that registered | 22 vs 33 | NIST IR 8242, OpenSAT 2017 pilot |
| Last OpenSAT evaluation and its domain | OpenSAT20 (2020), English public safety communications only | NIST OpenSAT evaluation series page |
| Scoring collar, Fearless Steps Phase-3 vs Phase-2 | 0.25 s vs 0.5 s | Joglekar et al., Fearless Steps Phase-3, Interspeech 2021 |
| Scoring collar used in DIHARD III | None (“without the use of forgiveness collars”), overlapped speech scored | Third DIHARD Diarization Challenge paper, arXiv 2020 |
| Agreement between the 3 human raters who labeled each AVA-Speech clip | Fleiss’ kappa 0.74 (“substantial”) | Chaudhuri et al., AVA-Speech, Interspeech 2018 |
| Segment length used to score Silero and rival VADs | 31.25 ms, threshold tuned on a 17-hour multi-domain validation set | Silero VAD Quality Metrics wiki (project claim) |
Context: even the ground truth is noisy. An Amazon study of 450 hours of movies and TV found that high-quality subtitle files used as labels contain up to 15% label noise (Sharma et al., Neurocomputing 2022). Most recent available data for the NIST and AVA scoring rules: 2017 to 2019 and 2018.
2. Challenge Results: Miss and False Alarm Rates on Hard Audio
Public challenges are where VAD meets audio nobody tuned for. The DIHARD III numbers show the core problem: the baseline’s false alarm rate jumped from 3.98% on development data to 15.07% on the evaluation core set, while its miss rate moved far less. Detectors overfit to the background noise they were trained on, and new noise turns into phantom speech.
The Fearless Steps results show the opposite end. NASA Apollo mission audio is noisy, multi-channel and decades old, yet the best Phase-3 system brought detection cost down to 1.47%, about a tenth of the baseline. The Fearless Steps Phase-3 overview also shows the limit: Phase-3 did not beat the Phase-2 best of 1.07%, and the Apollo-13 subset stayed harder than Apollo-11.
| Metric | Value | Source |
|---|---|---|
| DIHARD III baseline, development core set: miss / false alarm / overall error | 1.84% / 3.98% / 2.30% | Third DIHARD Diarization Challenge paper, arXiv 2020 |
| DIHARD III baseline, evaluation core set: miss / false alarm / overall error | 4.97% / 15.07% / 7.26% | Third DIHARD Diarization Challenge paper, arXiv 2020 |
| DIHARD III baseline, evaluation full set: miss / false alarm / overall error | 4.35% / 14.65% / 6.51% | Third DIHARD Diarization Challenge paper, arXiv 2020 |
| DIHARD III evaluation core set size and speech share | 184 recordings, 22.73 hours, 77.35% speech, 11 domains | Third DIHARD Diarization Challenge paper, arXiv 2020 |
| Fearless Steps Phase-3 participation (all tracks) | 9 organizations in 14 task-specific teams, 235 system submissions | Joglekar et al., Interspeech 2021 |
| Fearless Steps Phase-3 baseline detection cost, development vs evaluation | 12.84% vs 15.16% | Joglekar et al., Interspeech 2021 |
| Fearless Steps Phase-3 best detection cost: full evaluation / Apollo-13 subset / unseen-channel subset | 1.47% / 2.37% / 1.67% | Joglekar et al., Interspeech 2021 |
| Best Fearless Steps Phase-2 detection cost (still the state of the art after Phase-3) | 1.07% | Joglekar et al., Interspeech 2021 |
Outlier note: the DIHARD III paper says speech activity detection “remains a challenging problem for some of the covered domains,” and that without accurate segmentation diarization degraded “catastrophically for some,” with audiobooks the exception. Full tables are in the DIHARD III challenge paper. Most recent available data: DIHARD III (2020 to 2021) and Fearless Steps Phase-3 (2021); neither challenge has run a newer speech activity track.
3. Open Models Head to Head: pyannote, Silero and WebRTC
Most products ship one of three open detectors: the signal-processing VAD from Google’s WebRTC project, the Silero neural VAD, or the pyannote segmentation model. The independent data agree on the ranking, and the gap is largest exactly where products fail: noise-only audio. On ESC-50 environmental noise clips, WebRTC VAD scored 0 accuracy, meaning it called every noise clip speech, while Silero v6 got 0.87 right (Silero VAD Quality Metrics, the project’s own claim).
The pyannote 2.0 paper is the better reference for the false alarm vs miss split, because it reports both separately on three public sets (Bredin and Laurent, arXiv 2021). Its tuning minimizes their sum, so it lands near a balance (3.9% vs 3.3% on DIHARD 3). The 2021-era Silero model in the same table shows the classic lopsided profile: few misses, many false alarms. ROC-AUC, used below, is a 0-to-1 score of how well a detector separates speech from non-speech across all thresholds; the Matthews correlation coefficient (MCC) is a -1 to 1 score that stays honest when one class dominates.
| Metric | Value | Source |
|---|---|---|
| DIHARD 3: false alarm / miss / total detection error, pyannote 2.0 vs Silero VAD (2021 version) | 3.9% / 3.3% / 7.3% vs 17.0% / 4.0% / 21.0% | Bredin and Laurent, arXiv 2021 |
| AMI meetings: false alarm / miss, pyannote 2.0 vs Silero VAD (2021 version) | 3.6% / 3.2% vs 9.4% / 1.7% | Bredin and Laurent, arXiv 2021 |
| VoxConverse: false alarm / miss, pyannote 2.0 vs Silero VAD (2021 version) | 1.8% / 0.8% vs 3.0% / 1.1% | Bredin and Laurent, arXiv 2021 |
| ROC-AUC on a 17-hour multi-domain validation set: WebRTC / Silero v5 / Silero v6 | 0.73 / 0.96 / 0.97 | Silero VAD Quality Metrics wiki (project claim) |
| ROC-AUC on 80 hours of noisy MSDWild speech: WebRTC vs Silero v6 | 0.62 vs 0.79 | Silero VAD Quality Metrics wiki (project claim) |
| Share of ESC-50 noise clips correctly rejected: WebRTC / Silero v5 / Silero v6 | 0 / 0.61 / 0.87 | Silero VAD Quality Metrics wiki (project claim) |
| Utterance-level accuracy of Silero VAD v5 on ESC-50 (independent check) | 61% | Wang et al., Calm-Whisper, arXiv 2025 |
| MCC at 50 ms windows on 1.1 hours of podcasts, movies and music: RMS / WebRTC / Silero | 0.11 / 0.41 / 0.72 | McKinnon et al., Google, arXiv 2026 |
Context: the Google team found that adding hysteresis (a rule that keeps the current speech or non-speech state until the evidence clearly changes) lifted WebRTC’s MCC from 0.41 to 0.47 and did nothing useful for Silero (McKinnon et al., arXiv 2026). Silero’s own v6.0 release notes claim 16% fewer errors on noisy real-life data and 11% fewer on multi-domain validation than v5, a project claim not yet reproduced independently.
4. Noise and SNR: Where Detection Breaks Down
Signal-to-noise ratio (SNR) is the level of speech above the background, in decibels; 0 dB means speech and noise are equally loud. Every dataset here shows the same pattern: clean speech is a solved problem, and music and noise are not. In Google’s AVA-Speech movie dataset, clean speech sat at an estimated 40.8 dB SNR, but speech over music at only 11.7 dB, and even a large ResNet detector’s true positive rate fell from 0.992 on clean speech to 0.787 on speech with music (Chaudhuri et al., AVA-Speech).
Fixing noisy speech recall usually costs false alarms. Huawei’s 2025 tiny VAD raised noisy-speech accuracy on MS-SNSD from 15.9% to 89.9% with a preprocessing stage, but its non-speech accuracy on the same set fell from 99.1% to 87.7% (Jafarzadeh Asl et al., arXiv 2025). There is no free lunch: the operating point just moves along the trade-off curve.
| Metric | Value | Source |
|---|---|---|
| AVA-Speech estimated SNR: clean speech / speech with noise / speech with music | 40.8 dB / 16.2 dB / 11.7 dB | Chaudhuri et al., AVA-Speech, Interspeech 2018 |
| AVA-Speech true positive rate at a fixed 0.315 false positive rate, WebRTC VAD: clean / noise / music | 0.786 / 0.706 / 0.733 | Chaudhuri et al., AVA-Speech, Interspeech 2018 |
| AVA-Speech true positive rate at the same false positive rate, ResNet-960 model: clean / noise / music | 0.992 / 0.944 / 0.787 | Chaudhuri et al., AVA-Speech, Interspeech 2018 |
| Average F1 and AUC at 0 dB (babble, car, machine noise), WebRTC VAD vs MagicNet | F1 0.581 vs 0.859; AUC 0.691 vs 0.956 | Jia et al., arXiv 2024 |
| MS-SNSD noisy-speech accuracy, SG-VAD baseline vs Huawei VAD with preprocessing | 15.9% vs 89.9% | Jafarzadeh Asl et al., arXiv 2025 |
| MS-SNSD non-speech accuracy, SG-VAD baseline vs Huawei VAD with preprocessing | 99.1% vs 87.7% | Jafarzadeh Asl et al., arXiv 2025 |
| False positive rate needed to catch 99% of speech, SG-VAD vs Huawei VAD with preprocessing | 58% vs 28% | Jafarzadeh Asl et al., arXiv 2025 |
| Personal (target-speaker) VAD mean average precision, baseline vs best model: clean test set vs test set with added noise and reverberation | 0.900 vs 0.969 clean; 0.801 vs 0.928 noisy | Ding et al., Google, Personal VAD, arXiv 2019 |
Context: a December 2025 study trained hybrid CNN detectors on Kazakh speech mixed with noise from -18 dB to +30 dB; the best models reached a 99.6% F1 score and stayed robust above -12 dB, while models trained at a single noise level generalized poorly (Kulakayeva et al., ETASR 2025). The Amazon movie-and-TV review also found that models of about 2M parameters were less robust to noise than models of about 0.5M. Most recent available data for AVA-Speech and Personal VAD: 2018 and 2019.
5. Frame Size, Latency and Compute Budget
VAD runs before everything else, so it has to be cheap. The budgets in the literature are tiny: WebRTC VAD decides on 10, 20 or 30 ms frames (py-webrtcvad documentation), and in the AVA-Speech benchmark it needed 0.06 ms per decision, against 265 ms for the most accurate ResNet model, which in return lifted the true positive rate on all speech from 0.722 to 0.917.
The 2024 to 2026 work closes that gap from the model side. MagicNet runs at a real-time factor (processing time divided by audio duration) of 0.034 with 22.7K parameters, close to WebRTC’s 0.015, while Silero says its roughly 2 MB model handles a 30+ ms chunk in under 1 ms on one CPU thread (Silero VAD README, project claim). Personal VAD shows the other reason to stay small: Google built it to gate on-device speech recognition so the recognizer only wakes for the enrolled user, cutting compute and battery use.
| Metric | Value | Source |
|---|---|---|
| WebRTC VAD frame lengths, aggressiveness modes and sample rates | 10, 20 or 30 ms; modes 0 to 3; 8, 16, 32 or 48 kHz, 16-bit mono PCM | py-webrtcvad documentation |
| Latency per decision, AVA-Speech benchmark: WebRTC / CNN tiny_320 / CNN ResNet-960 | 0.06 ms / 2.23 ms / 265 ms | Chaudhuri et al., AVA-Speech, Interspeech 2018 |
| Real-time factor on an Intel Xeon CPU: WebRTC / MagicNet / MarbleNet | 0.015 / 0.034 / 0.038 | Jia et al., arXiv 2024 |
| Parameters: MagicNet / CRNN / MarbleNet | 22.7K / 36.8K / 85.2K | Jia et al., arXiv 2024 |
| Silero VAD processing time and model size | Under 1 ms per 30+ ms chunk on one CPU thread; about 2 MB; ONNX up to 4-5x faster | Silero VAD README (project claim) |
| pyannote 2.0 segmentation model frame step and input chunk | About 16 ms; 5 s chunks | Bredin and Laurent, arXiv 2021 |
| EML online SAD decision interval and real-time factor on one CPU (Apollo audio) | About every 0.1 s; about 0.002 | Ghahabi and Fischer, arXiv 2021 |
| Personal VAD size, embedding-only model vs speaker-verification-based pipeline | 0.13M vs 4.88M + 0.06M parameters, one decision every 10 ms | Ding et al., Google, Personal VAD, arXiv 2019 |
Context: frame size is also an accuracy lever. The Google 2026 window study averaged predictions over windows from 10 ms up to 10 s, and found Silero clearly ahead of WebRTC at the 100 ms setting. Longer windows smooth out flicker but add delay, which is why real-time voice tools keep the native frame short and smooth afterward.
6. Telephony DTX: Bandwidth Saved by Not Sending Silence
In telephony, VAD pays for itself through discontinuous transmission (DTX): when the detector hears no speech, the phone stops sending voice packets and the far end plays synthetic “comfort noise” so the line does not sound dead. Avaya’s documentation notes that more than 40% of a conversation’s transmission is silence, because usually only one party speaks at a time (Avaya Aura documentation, vendor guidance).
The planning numbers are more cautious than the silence share suggests. Cisco tells engineers to count on VAD savings only across many simultaneous calls (Cisco per-call bandwidth guide, vendor guidance), because a single call can be all talk, and music on hold or fax defeat the detector. Avaya names the error trade-off directly: too aggressive and the beginnings and ends of words are clipped (misses), not aggressive enough and no bandwidth is saved (false alarms). For how packet timing then shapes call quality, see our VoIP jitter buffer and POLQA statistics.
| Metric | Value | Source |
|---|---|---|
| Bandwidth saving from VAD, as an average over more than 24 calls | Up to 35 percent, not realized on every individual call | Cisco, Voice Over IP Per Call Bandwidth Consumption (vendor guidance) |
| Links on which Cisco says VAD savings must not be counted | Links carrying fewer than 24 simultaneous voice calls | Cisco, Voice Over IP Per Call Bandwidth Consumption (vendor guidance) |
| Per-call Ethernet bandwidth at default payload sizes, G.711 vs G.729 | 87.2 kbps vs 31.2 kbps | Cisco, Voice Over IP Per Call Bandwidth Consumption (vendor guidance) |
| Share of a conversation’s transmission that is silence | More than 40% | Avaya Aura 10.2.x documentation (vendor guidance) |
| Default state of silence suppression in Avaya Communication Manager | Disabled, except for G.729B where it is built in | Avaya Aura 10.2.x documentation (vendor guidance) |
| ITU-T G.729 core bit rate; Annex B role | 8 kbit/s; Annex B adds VAD, DTX and comfort noise, tuned for circuits, with Appendices II and III for VoIP | ITU-T Rec. G.729 (06/2012) summary |
| EVS codec comfort-noise update interval: default / fixed-rate range / adaptive range | Every 8 frames / 1 to 100 frames / 8 to 50 frames | 3GPP TS 26.449, EVS comfort noise generation |
Context: 3GPP’s EVS overview says DTX reduces the average required bit-rate, “leading to a more efficient transmission with decreased load and hence increased capacity,” but gives no single percentage (3GPP TS 26.441 overview). The ITU-T G.729 summary dates from June 2012; it remains the most recent available text for that codec family.
7. Downstream Damage: How VAD Errors Propagate into ASR and Diarization
A VAD error is rarely the final error. A false alarm hands the speech recognizer noise to transcribe; a miss deletes words no later stage can recover. The clearest controlled result comes from AGH University of Krakow: on Common Voice speech padded with non-speech sounds at 9 dB SNR, Whisper’s word error rate was 68.3% with WebRTC VAD in front of it and 8.0% with Silero VAD (Barański et al., arXiv 2025). The detector choice alone moved the error by a factor of about eight. Rates above 100% are possible here because invented words count as insertions. Our speech recognition hallucination statistics cover the recognizer side of this failure in depth.
Good VAD also buys speed and timing accuracy. WhisperX uses pyannote VAD to cut audio into speech-only chunks and batch them, which made it 11.8 times faster than plain Whisper and raised word-timing precision on AMI meetings (Bain et al., WhisperX, arXiv 2023). On the diarization side, a 2025 ETH Zurich benchmark across 196.6 hours found that “missed speech constitutes the dominant failure case across all diarization models” (Lanzendörfer et al., arXiv 2025); our speaker diarization error rate statistics track the full error budget.
| Metric | Value | Source |
|---|---|---|
| Whisper word error rate on speech with non-speech sounds before or after it, WebRTC VAD vs Silero VAD | 68.3% vs 8.0% | Barański et al., arXiv 2025 |
| Same comparison when the noise overlaps the speech, WebRTC VAD vs Silero VAD | 75.4% vs 10.8% | Barański et al., arXiv 2025 |
| Transcription speed with VAD-based batching, WhisperX vs Whisper | 11.8x vs 1.0x | Bain et al., WhisperX, arXiv 2023 |
| Word timing precision / recall on AMI meetings (200 ms collar), WhisperX vs Whisper | 84.1 / 60.3 vs 78.9 / 52.1 | Bain et al., WhisperX, arXiv 2023 |
| Average non-vocal time per recording, speakers with aphasia vs control speakers (measured with pyannote, confirmed with Silero) | 6.8 s vs 1.3 s (p < 2.2e-16) | Koenecke et al., Careless Whisper, ACM FAccT 2024 |
| Average non-vocal time, recordings where Whisper hallucinated vs recordings where it did not | 5.3 s vs 3.5 s | Koenecke et al., Careless Whisper, ACM FAccT 2024 |
| Extra diarization difficulty when systems must find speech themselves (Track 2) instead of receiving reference speech regions (Track 1) | “On the order of 5%” for both DER and JER | Third DIHARD Diarization Challenge paper, arXiv 2020 |
| Diarization benchmark scope behind the missed-speech finding | 196.6 hours, 4 datasets, 5 languages | Lanzendörfer et al., ETH Zurich, arXiv 2025 |
Outlier note: the same Krakow table shows that VAD is necessary but not sufficient. Silero VAD plus a de-looping step and a list of known hallucinated phrases did better than Silero alone (6.5% vs 8.0% word error rate on non-overlapped audio). Most recent available data for WhisperX: 2023.
Summary: Voice Activity Detection by the Numbers
| Metric | Value | Source |
|---|---|---|
| NIST speech activity cost function | 0.75 x miss rate + 0.25 x false alarm rate | NIST IR 8242, OpenSAT 2017 pilot |
| DIHARD III baseline, evaluation core: miss / false alarm | 4.97% / 15.07% | Third DIHARD Diarization Challenge paper, arXiv 2020 |
| Fearless Steps Phase-3 detection cost, baseline vs best (evaluation) | 15.16% vs 1.47% | Joglekar et al., Interspeech 2021 |
| Fearless Steps scoring collar, Phase-3 vs Phase-2 | 0.25 s vs 0.5 s | Joglekar et al., Interspeech 2021 |
| DIHARD 3 false alarm / miss, pyannote 2.0 | 3.9% / 3.3% | Bredin and Laurent, arXiv 2021 |
| DIHARD 3 false alarm / miss, Silero VAD (2021 version) | 17.0% / 4.0% | Bredin and Laurent, arXiv 2021 |
| ESC-50 noise clips correctly rejected, WebRTC vs Silero v6 | 0 vs 0.87 | Silero VAD Quality Metrics wiki (project claim) |
| MCC at 50 ms windows, WebRTC vs Silero | 0.41 vs 0.72 | McKinnon et al., Google, arXiv 2026 |
| Average AUC at 0 dB SNR, WebRTC vs MagicNet | 0.691 vs 0.956 | Jia et al., arXiv 2024 |
| AVA-Speech ResNet true positive rate, clean vs speech with music | 0.992 vs 0.787 | Chaudhuri et al., AVA-Speech, Interspeech 2018 |
| MS-SNSD noisy-speech accuracy, baseline vs Huawei VAD with preprocessing | 15.9% vs 89.9% | Jafarzadeh Asl et al., arXiv 2025 |
| Latency per decision, WebRTC vs ResNet-960 | 0.06 ms vs 265 ms | Chaudhuri et al., AVA-Speech, Interspeech 2018 |
| VAD bandwidth saving averaged over more than 24 calls | Up to 35 percent | Cisco, Per Call Bandwidth Consumption (vendor guidance) |
| Share of a conversation’s transmission that is silence | More than 40% | Avaya Aura 10.2.x documentation (vendor guidance) |
| EVS default comfort-noise update interval | Every 8 frames | 3GPP TS 26.449 |
| Whisper word error rate behind WebRTC VAD vs Silero VAD | 68.3% vs 8.0% | Barański et al., arXiv 2025 |
| WhisperX speed-up from VAD-based batching | 11.8x | Bain et al., WhisperX, arXiv 2023 |
Methodology and Sources
- NIST IR 8242, OpenSAT 2017 pilot evaluation report: cost function, collar, participation.
- NIST OpenSAT evaluation series page: evaluation years, domains and program status.
- The Third DIHARD Diarization Challenge paper (arXiv 2012.01477): baseline speech activity detection miss and false alarm rates, dataset sizes, Track 1 vs Track 2 findings.
- Joglekar et al., Fearless Steps Challenge Phase-3, Interspeech 2021: Apollo audio detection cost results and collars.
- Ghahabi and Fischer, EML online SAD for Fearless Steps Phase-III, arXiv 2021: decision interval and real-time factor.
- Bredin and Laurent, End-to-end speaker segmentation for overlap-aware resegmentation, arXiv 2021: false alarm and miss rates on AMI, DIHARD 3 and VoxConverse.
- Silero VAD Quality Metrics wiki, Silero VAD README and Silero VAD releases: project-reported accuracy, speed and release claims.
- McKinnon et al., Window Size Versus Accuracy Experiments in Voice Activity Detectors, Google, arXiv 2026: MCC comparison of RMS, WebRTC and Silero.
- py-webrtcvad documentation: WebRTC VAD frame sizes, modes and sample rates.
- Chaudhuri et al., AVA-Speech, Interspeech 2018: label shares, estimated SNR, per-condition true positive rates and latency.
- Jia et al., A Real-Time Voice Activity Detection Based on Lightweight Neural Network, arXiv 2024: 0 dB F1, AUC, parameters and real-time factor.
- Jafarzadeh Asl et al., Tiny Noise-Robust Voice Activity Detector for Voice Assistants, Huawei, arXiv 2025: noisy-speech vs non-speech accuracy trade-off.
- Kulakayeva et al., Robust Neural Network-Based VAD under Low SNR, ETASR 2025: SNR range and robustness threshold.
- Sharma et al., Empirical review of VAD for movies and TV shows, Neurocomputing 2022: label noise and model size findings.
- Ding et al., Personal VAD, Google, arXiv 2019: target-speaker VAD precision, size and frame rate.
- Cisco, Voice Over IP Per Call Bandwidth Consumption: VAD savings guidance and per-call bandwidth.
- Avaya Aura 10.2.x, Silence suppression or voice activity detection: silence share and defaults.
- ITU-T Rec. G.729 (06/2012) summary, 3GPP TS 26.449 EVS comfort noise generation and 3GPP TS 26.441 EVS overview: DTX and comfort noise mechanics (3GPP text read through the iTecSpec mirror).
- Barański et al., Investigation of Whisper ASR Hallucinations Induced by Non-Speech Audio, arXiv 2025: word error rate behind WebRTC vs Silero VAD.
- Bain et al., WhisperX, arXiv 2023: speed-up and word timing results from VAD-based batching.
- Wang et al., Calm-Whisper, arXiv 2025: independent check of Silero v5 accuracy on ESC-50.
- Koenecke et al., Careless Whisper, ACM FAccT 2024: non-vocal time and hallucination.
- Lanzendörfer et al., Benchmarking Diarization Models, ETH Zurich, arXiv 2025: missed speech as the dominant diarization failure.
- Data watch: much of the benchmark base is older than three years (OpenSAT 2017, AVA-Speech 2018, Personal VAD 2019, DIHARD III 2020 to 2021, Fearless Steps Phase-3 2021, WhisperX 2023), because no public speech activity challenge has run since; treat those as the most recent available data. The DIHARD 3 baseline row in the pyannote paper (4.0% false alarm, 4.2% miss) does not match the DIHARD III paper’s own evaluation numbers (15.07% and 4.97% on the core set), probably because of a different partition or setup, so we cite each only with its own source. The Silero model in the 2021 pyannote table predates Silero v5 and v6. The Google 2026 paper gives Silero a 16 ms native window while the Silero README speaks of 30+ ms chunks. AVA-Speech describes itself as both “45 hours” and “about 46 hours.” The Personal VAD paper does not fully say whether the noisy-test models were trained differently. Silero, Cisco and Avaya figures are the publishers’ own claims or guidance. Cisco’s 35 percent is an average-case planning figure, not a measurement of any one network.
- Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when the ICASSP 2027 papers are published and the Silero VAD project next updates its Quality Metrics page.