Researchers observed smart speakers waking up by mistake 0.95 times per hour in a controlled test that played two rounds of 134 hours of Netflix content to seven devices (Northeastern University and Imperial College London, PETS 2020). A separate lab study that ran 11 speakers from 8 manufacturers through TV shows and newscasts released more than 1,000 verified accidental triggers (Ruhr University Bochum and Max Planck Institute, arXiv 2020). Those false accepts now carry a legal price: Google’s $68 million settlement over Assistant false accepts received 1.49 million timely claims before final approval in October 2026 (U.S. District Court N.D. Cal., Kumandan v. Google order 2026). On the recognition side, a home dinner-party benchmark recorded with 7-microphone arrays saw word error rate jump from 42.10% on close-talk headsets to 77.52% on the far-field devices (DiPCo corpus paper, 2019). We aggregated data from Northeastern University, Ruhr University Bochum, Amazon Science, Apple Machine Learning Research, Google Research, the CHiME challenge organizers, Johns Hopkins University and two federal class-action records. How many homes own these devices is covered in our smart speaker usage statistics; this roundup is about how well they hear.
TL;DR
- Researchers observed 0.95 smart speaker misactivations per hour, or 1.43 per 10,000 spoken words, while TV shows played nearby (Northeastern University and Imperial College London, PETS 2020).
- The Cortana wake word produced 197 accidental triggers vs 8 for Hey Siri in 624 hours of audio per speaker (Ruhr University Bochum and Max Planck Institute, arXiv 2020).
- Smart speaker users who noticed an accidental wake at least once a month: 43.7% (Voicebot.ai survey 2018, reported by Voicebot.ai 2020).
- A second-stage classifier cut Alexa false alarms by about 67% on top of the first-stage detector (Amazon Science, ICASSP 2018).
- Apple’s streaming transformer suppresses 95% of false triggers with one extra second of post-trigger audio (Apple Machine Learning Research 2021).
- HomePod’s six-microphone processing improved word error rate by about 90% relative during music playback (Apple Machine Learning Research 2018).
- Wake word false rejects at one false alarm per hour: 2.00% close-talk vs 5.11% on far-field arrays (Personalized Voice Trigger Challenge 2020).
- Word error rate on a home dinner-party corpus: 42.10% close-talk vs 77.52% far-field (DiPCo corpus paper, 2019).
- Even the best systems in the CHiME-7 and CHiME-8 challenges stay above 30% tcpWER on the CHiME-6 home dinner parties (CHiME-7/8 DASR review, arXiv 2025).
- 24.76% of 2,552 Alexa interactions by adults aged 66 to 94 involved an error (Johns Hopkins University, arXiv 2024).
- Google’s $68 million Assistant false accept settlement received 1.49 million timely claims (Kumandan v. Google final approval order, October 2026).
- Apple’s $95 million Siri settlement pays up to $20 per device for unintended activations (Lopez v. Apple settlement administrator, 2025-2026).
1. Accidental Activations: How Often TV Dialogue Wakes a Speaker
0.95 misactivations per hour sounds small until you remember that a television in a living room runs for hours at a time, and that the same study found almost all activations detected locally (the device lit up) were also sent to the cloud. The privacy question is therefore not only how often a speaker wakes, but how long it keeps listening once it does.
Duration is where the Northeastern and Imperial College study is most revealing: the longest tenth of activations ran 8 to 10 seconds depending on the device, long enough to capture a full sentence of private conversation. Misactivations were also not random noise: 20.7% of Google Home Mini activations and 17.7% of HomePod activations repeated in more than 75% of the 12 test runs, which means specific lines of dialogue reliably fool specific devices. Owner surveys, reported by Voicebot.ai, tell the same story from the couch.
| Metric | Value | Source |
|---|---|---|
| Content played to the test speakers | Two rounds of 134 hours from 12 TV shows, in the US and the UK, to 7 devices (Google Home Mini, Apple HomePod, Harman Kardon Invoke, 2 Echo Dot 2nd gen, 2 Echo Dot 3rd gen) | Northeastern University and Imperial College London, PETS 2020 |
| Misactivation rate observed | 0.95 per hour, or 1.43 per 10,000 words spoken | Northeastern University and Imperial College London, PETS 2020 |
| TV shows that caused the most activations | The West Wing: 4.26 average activations per hour; Narcos: 6.21 activations per 10,000 words of dialogue | Northeastern University and Imperial College London, PETS 2020 |
| Length reached by the longest 10% of activations; longest single activation | At least 10 seconds on HomePod, 9 seconds on Google Home Mini, 8 seconds on Echo Dot 2nd gen (Echo wake word); up to 43 seconds in one experiment, treated as an outlier | Northeastern University and Imperial College London, PETS 2020 |
| Smart speaker users who noticed they accidentally woke their voice assistant at least once a month | 43.7% | Voicebot.ai survey 2018, reported by Voicebot.ai 2020 |
| US voice search users (phones included) who accidentally woke their assistant over a month | 64% of 501 respondents | The Manifest survey, reported by Voicebot.ai 2020 |
| Share of reported accidental wakes where the software detected its wake word for no apparent reason | 29% | The Manifest survey, reported by Voicebot.ai 2020 |
Most recent available data: Northeastern University and Imperial College London, PETS 2020, and owner surveys from 2018 to 2020. No larger independent misactivation measurement has been published since, so these remain the reference figures. The widely repeated headline “up to 19 times per day” came from press coverage, not from the study’s own reported rates.
2. Which Wake Words Misfire Most
The gap between wake words is enormous: Cortana produced 197 accidental triggers while Hey Siri produced 8 under the same playback. Short, common-sounding wake words collide with everyday speech; the researchers found the Chinese word ‘dage’ (big brother) confused with ‘Echo’ and the German phrase ‘Am Sonntag’ confused with ‘Amazon’ in weather forecasts.
The cloud acts as a second filter, but an imperfect one. Only a small share of Cortana’s local triggers survived the server check, yet the accidental triggers paper found that a trigger confirmed in the cloud can upload 10 seconds or more of audio. A cloud verification step, in plain words, is a second and larger recognizer on the company’s servers that re-listens to the clip and cancels the wake-up if it disagrees with the device.
| Metric | Value | Source |
|---|---|---|
| Test scope | 11 smart speakers from 8 manufacturers; 13 main datasets of 24 hours each, played twice (13 x 24 x 2 = 624 hours per speaker) | Ruhr University Bochum and Max Planck Institute, arXiv 2020 |
| Accidental triggers by wake word | Cortana 197; Alexa 99; Computer 77; Amazon 57; Echo 43; OK/Hey Google 16; Hey Siri 8 | Ruhr University Bochum and Max Planck Institute, arXiv 2020 |
| Cortana triggers that also passed the cloud-side check | 8 of 197 (8 / 197 = 4.1%) | Ruhr University Bochum and Max Planck Institute, arXiv 2020 |
| Amazon and Google triggers that fired again at least 4 times in 10 replays | About 75% | Ruhr University Bochum and Max Planck Institute, arXiv 2020 |
| ’Computer’ wake word triggered by the same Modern Family episodes, English version vs German version | 9 times vs 1 time | Ruhr University Bochum and Max Planck Institute, arXiv 2020 |
| Audio uploaded after a cloud-confirmed trigger vs time for a local-only trigger to stop | 10 seconds or more vs about 1-2 seconds | Ruhr University Bochum and Max Planck Institute, arXiv 2020 |
Most recent available data: Ruhr University Bochum and Max Planck Institute, arXiv 2020. The count for Alexa appears as 99 in the paper’s text and as 100 in a summary table row; we use the text figure.
3. Wake Word Detection: Trading False Rejects for False Alarms
Every wake word engine balances two errors. A false reject is when the speaker ignores a real ‘Alexa’ or ‘Hey Siri’; a false alarm (also called a false accept or false trigger) is when it wakes although nobody said the wake word. Engineers fix one error rate and then try to push the other down, which is why results are quoted as “relative reduction at a fixed false alarm level” or “at 0.1 false accepts per hour”.
The pattern across vendors is a layered defence. The headline result here is Apple’s: 95% of false triggers suppressed with just one extra second of audio (Apple Machine Learning Research 2021). Amazon’s two-stage detector and Apple’s lattice-based checks show the same logic: a cheap always-on detector wakes the device, and a heavier second stage decides whether to keep listening.
| Metric | Value | Source |
|---|---|---|
| Two-stage on-device detector with background speech modeling: false rejects at a fixed false-alarm level; false alarms at a fixed miss rate | About 16% relative reduction; about 37% relative reduction | Amazon Science, ICASSP 2018 |
| Second-stage classifier on top of the first-stage detector | About 67% fewer false alarms | Amazon Science, ICASSP 2018 |
| False trigger rate with a complementary language model, at a fixed 0.4% false suppression of real requests | 38.34% relative reduction, then a further 10.8% with a parallel model | Apple, Interspeech 2020 |
| False triggers suppressed by a streaming transformer with one second of post-trigger audio | 95% | Apple Machine Learning Research 2021 |
| Same streaming model vs the non-streaming version | 18% average relative reduction in false rejects; 32% less runtime memory; 56% less inference time | Apple Machine Learning Research 2021 |
| Multi-microphone network vs a single-channel model run on each microphone, at 0.1 false accepts per hour | 27.2% (clean) and 31.8% (noisy) relative reduction in false rejects on dual-microphone test sets | Google Research, Odyssey 2020 |
| Siri voice trigger design | Two trigger phrases (‘Hey Siri’ and ‘Siri’) across 7 device categories; first-pass network compressed to 4 bits per weight | Apple Machine Learning Research 2023 |
Context: vendors almost never publish absolute false-accept rates for shipping products, only relative improvements. The Complementary Language Model paper and the Google multi-channel paper follow that convention. Most recent available data for the shipping-system architecture: Apple Machine Learning Research, Voice Trigger System for Siri, 2023.
4. Beamforming and the Acoustic Front End
Beamforming means combining the signals from several microphones so that the array ‘listens’ toward the talker and away from noise. Echo cancellation means subtracting the speaker’s own music from what its microphones hear. Both matter because, as Apple measured, the device’s own playback can be 30-40 dB louder than the voice it is trying to hear (Apple Machine Learning Research 2018).
The payoff is largest exactly where a smart speaker struggles most: during its own music playback. Amazon’s layered echo cancellation work reports similar orders of magnitude. The same directional principle shrinks noise at the ear in hearing aids, as our directional microphone hearing aid SNR statistics show; a smart speaker applies it from across the room.
| Metric | Value | Source |
|---|---|---|
| Microphones on Apple HomePod; processing cost of its speech enhancement | Six; less than 15% of a single A8 core at 1.4 GHz | Apple Machine Learning Research 2018 |
| Device playback (echo) vs far-field speech; residual echo after linear echo cancellation at loud volume | 30-40 dB louder; 10-20 dB above the far-field speech | Apple Machine Learning Research 2018 |
| Relative word error rate improvement from HomePod multichannel processing | About 40% (reverberant speech), 90% (playback), 74% (loud background noise), 61% (competing talker) | Apple Machine Learning Research 2018 |
| Absolute improvement in ‘Hey Siri’ false rejects with a competing talker | 29.0% vs mask-based noise reduction; 30.3% vs the baseline signal processing output | Apple Machine Learning Research 2018 |
| Echo and noise reduction from Amazon’s layered system on Echo devices | About 40 dB echo reduction; about 19 dB noise reduction | Amazon Science 2018 |
| Relative word error rate improvement of the three layers (labelled SBNR, SBRES and NLEC by Amazon) | About 22%, 14% and 17% | Amazon Science 2018 |
| Far-field device array in the DiPCo home recordings | 7 microphones: 6 on a circle of 35 mm radius plus 1 in the centre | DiPCo corpus paper, 2019 |
| Arrays used in each CHiME-6 home dinner party | 6 Microsoft Kinect devices with 4-channel linear arrays (24 array microphones) | CHiME-6 Challenge paper, 2020 |
Most recent available data: Apple Machine Learning Research 2018 and Amazon Science 2018. Both companies have shipped newer hardware since, but neither has published comparable front-end measurements.
5. The Distance Penalty in Far-Field Recognition
Moving the microphone from the mouth to the other side of the room nearly doubles recognition errors in controlled tests. Word error rate rose from 42.10% on close-talk headsets to 77.52% on far-field devices in the same dinner-party sessions (DiPCo corpus paper, 2019), and wake word false rejects in the Personalized Voice Trigger Challenge were more than twice as high on far-field arrays as on close-talk phones.
Conversational home audio is far harder than a single command: people talk over each other, move between rooms and cook. Two terms help read the table. Guided source separation (GSS) is a multi-microphone method that uses who-spoke-when timing to pull each voice apart, and tcpWER is a word error rate that also penalises words assigned to the wrong speaker or the wrong moment. The CHiME-7 and CHiME-8 review shows that years of progress still leave home dinner parties as the hardest scenario.
| Metric | Value | Source |
|---|---|---|
| Wake word false rejects at one false alarm per hour, close-talk phone vs far-field microphone array (development set) | 2.00% vs 5.11% | Personalized Voice Trigger Challenge 2020 |
| Challenge recording setup | 612 hours, 658,995 utterances, 550 speakers; arrays of 4 or 6 channels at 1 m, 3 m and 5 m in a room built as a smart home | Personalized Voice Trigger Challenge 2020 |
| Far-field smart-home wake word (‘Xiao T, Xiao T’), evaluation set: false rejects and false alarms | Best audio-only system 6.38% and 4.65%; best audio-visual system 2.15% and 3.44%; official audio-visual baseline 10.1% and 15% | Duke Kunshan University system, ICASSP 2023 (MISP 2021 challenge) |
| Word error rate, close-talk headset vs far-field 7-microphone device (evaluation set) | 42.10% vs 77.52% | DiPCo corpus paper, 2019 |
| CHiME-6 baseline word error rate (development / evaluation), with GSS vs with a conventional beamformer (BeamformIt) | 51.8% / 51.3% vs 69.8% / 61.2% | CHiME-6 Challenge paper, 2020 |
| CHiME-6 baseline when the system must also find who spoke when (development / evaluation) | 84.3% / 77.9% | CHiME-6 Challenge paper, 2020 |
| Best CHiME-5 system word error rate (development / evaluation) | 45.6% / 46.6% (USTC-iFlytek) | CHiME-6 Challenge paper, 2020 |
| CHiME-6 tcpWER of the best CHiME-7/8 systems; CHiME-8 participation | Above 30%; 9 teams submitting 32 systems | CHiME-7/8 DASR review, arXiv 2025 |
Context: the challenge figures measure open conversation, not short commands, so they overstate the error a speaker makes on ‘set a timer’. They are still the best public evidence of how much the room itself costs, and the CHiME-6 paper shows the jump from a conventional beamformer to GSS alone was worth about 18 points on the development set (69.8% minus 51.8%).
6. Who Gets Missed: Older Adults at Home
Wake word and recognition failures are not spread evenly. In a month-long in-home deployment, 24.76% of Alexa interactions by adults aged 66 to 94 involved an error (Johns Hopkins University, arXiv 2024), and error rates did not fall over the four weeks. Hearing the request was only part of the problem: misunderstanding intent was the single largest error category.
The wake word itself trips people up. Saying a near-miss such as ‘Alexis’ accounted for more wrong turns than false activations did, which is the mirror image of the TV problem in section 1: the device is strict with its owner and lenient with the television. Gaps by accent and dialect are covered in our speech recognition accent bias statistics; this section stays with age and cognition.
| Metric | Value | Source |
|---|---|---|
| Interactions with an error, 15 households aged 66 to 94, Echo Dot, four weeks | 24.76% (632 of 2,552) | Johns Hopkins University, arXiv 2024 |
| Speech recognition errors as a share of all errors | 24.4% (154 of 632) | Johns Hopkins University, arXiv 2024 |
| Errors where Alexa did not pick up the request at all (‘not listened’) | 10.9% (69 of 632) | Johns Hopkins University, arXiv 2024 |
| Errors caused by saying the wrong wake word (for example ‘Alexis’) | 7.8% (49 of 632) | Johns Hopkins University, arXiv 2024 |
| Errors caused by false activations (mis-triggers); share resolved on retry | 2.4% (15 of 632); 6.7% | Johns Hopkins University, arXiv 2024 |
| Intent recognition failures, the largest error category | 32.3% (204 of 632) | Johns Hopkins University, arXiv 2024 |
| Speech recognition errors on Alexa commands read by 83 older adults, dementia group vs cognitively normal group | Significantly higher in the dementia group | Cohn et al., arXiv 2026 |
The 2026 dementia study reports that speaker intensity, voice quality and pause ratio predicted recognition accuracy, which links the age gap back to acoustics: quieter, breathier voices lose more energy across a room before they reach the array.
7. False Accepts in Court and in Company Policy
False accepts have moved from lab metric to legal liability. Google’s settlement has a class period that starts on May 18, 2016 and a privacy class defined around false accept recordings, and 1.49 million timely claims arrived against notices emailed to 173,814,291 addresses (U.S. District Court N.D. Cal., Kumandan v. Google final approval order, October 2026). By our calculation that is a claim rate of about 0.86% of notified addresses (1.49 million / 173,814,291); the order itself states no claim rate.
Apple’s parallel Siri settlement covers unintended Siri activations during private conversations over a decade. Company policy has moved too: Apple stated in January 2025 that it does not retain audio recordings of Siri interactions unless users explicitly opt in (Apple Newsroom 2025).
| Metric | Value | Source |
|---|---|---|
| Google Assistant false accept settlement fund; final approval date | $68,000,000, non-reversionary; October 13, 2026 | U.S. District Court N.D. Cal., Kumandan v. Google order 2026 |
| Notice delivery and opt-outs | Emails to 173,814,291 addresses, 98.76% delivered; 217 valid exclusion requests and 7 objections | U.S. District Court N.D. Cal., Kumandan v. Google order 2026 |
| Estimated payment per class member | About $18 to $56 (purchaser class); about $2 to $10 (privacy class) | U.S. District Court N.D. Cal., Kumandan v. Google order 2026 |
| Attorneys’ fees awarded | $22,666,666.70 (one third of the fund) | U.S. District Court N.D. Cal., Kumandan v. Google order 2026 |
| Apple Siri unintended activation settlement | $95 million fund; up to $20 per Siri device, up to five devices; class period September 17, 2014 to December 31, 2024 | Lopez v. Apple settlement administrator, 2025-2026 |
| Share of Assistant audio snippets reviewed by language experts | Around 0.2% | Google, The Keyword blog, July 2019 |
| Share of Siri requests heard by human graders before Apple suspended grading | Less than 0.2% | Apple Newsroom, August 2019 |
Context: Google’s 2019 statement on speech data described false accepts as rare, and Apple’s 2019 statement committed to deleting recordings judged to be inadvertent triggers. Neither company published a false accept frequency, which is why the lab studies in sections 1 and 2 remain the main public evidence.
Summary: Far-Field Smart Speaker Beamforming and Wake Word Accuracy by the Numbers
| Metric | Value | Source |
|---|---|---|
| Misactivations while TV shows played nearby | 0.95 per hour, or 1.43 per 10,000 words | Northeastern University and Imperial College London, PETS 2020 |
| Longest 10% of activations | 8 to 10 seconds or more, depending on device | Northeastern University and Imperial College London, PETS 2020 |
| Accidental triggers, Cortana vs Hey Siri, 624 hours per speaker | 197 vs 8 | Ruhr University Bochum and Max Planck Institute, arXiv 2020 |
| Smart speaker users noticing an accidental wake at least monthly | 43.7% | Voicebot.ai survey 2018, reported by Voicebot.ai 2020 |
| False alarms removed by a second-stage classifier | About 67% | Amazon Science, ICASSP 2018 |
| False trigger reduction from a complementary language model | 38.34% relative | Apple, Interspeech 2020 |
| False triggers suppressed with one extra second of audio | 95% | Apple Machine Learning Research 2021 |
| False reject reduction from a dual-microphone network at 0.1 false accepts per hour | 27.2% clean, 31.8% noisy | Google Research, Odyssey 2020 |
| Device playback vs far-field speech level | 30-40 dB louder | Apple Machine Learning Research 2018 |
| Word error rate improvement from HomePod processing during playback | About 90% relative | Apple Machine Learning Research 2018 |
| Echo reduction on Echo devices | About 40 dB | Amazon Science 2018 |
| Wake word false rejects, close-talk vs far-field | 2.00% vs 5.11% | Personalized Voice Trigger Challenge 2020 |
| Best audio-visual far-field wake word system, false rejects and false alarms | 2.15% and 3.44% | Duke Kunshan University system, ICASSP 2023 |
| Word error rate, close-talk vs far-field | 42.10% vs 77.52% | DiPCo corpus paper, 2019 |
| CHiME-6 baseline, GSS vs conventional beamformer (development set) | 51.8% vs 69.8% | CHiME-6 Challenge paper, 2020 |
| Best CHiME-7/8 systems on CHiME-6 home dinner parties | Above 30% tcpWER | CHiME-7/8 DASR review, arXiv 2025 |
| Alexa interactions with an error, adults aged 66 to 94 | 24.76% | Johns Hopkins University, arXiv 2024 |
| Google Assistant false accept settlement and timely claims | $68 million; 1.49 million | U.S. District Court N.D. Cal., Kumandan v. Google order 2026 |
| Apple Siri unintended activation settlement | $95 million; up to $20 per device | Lopez v. Apple settlement administrator, 2025-2026 |
Methodology and Sources
We ran 15 searches and read the primary paper, court record or company page for every figure above. Where only a relative improvement is published, we report it as relative; we did not convert relative improvements into absolute rates. The two settlement figures are court and administrator records, not technical measurements.
- Northeastern University and Imperial College London, When Speakers Are All Ears: Characterizing Misactivations of IoT Smart Speakers, PETS 2020
- Voicebot.ai, report on The Manifest survey and the Voicebot 2018 survey, January 2020
- Ruhr University Bochum and Max Planck Institute for Security and Privacy, Unacceptable, where is my privacy? Exploring Accidental Triggers of Smart Speakers, arXiv 2020
- Amazon Science, Monophone-based Background Modeling for Two-stage On-device Wake Word Detection, ICASSP 2018
- Amazon Science, multilayer echo cancellation and voice enhancement for smart speakers, 2018
- Apple Machine Learning Research, Complementary Language Model and Parallel Bi-LRNN for False Trigger Mitigation, Interspeech 2020
- Apple Machine Learning Research, Streaming Transformer for Hardware Efficient Voice Trigger Detection and False Trigger Mitigation, 2021
- Apple Machine Learning Research, Voice Trigger System for Siri, 2023
- Apple Machine Learning Research, Optimizing Siri on HomePod in Far-Field Settings, 2018
- Google Research, Small Footprint Multi-channel Keyword Spotting, Odyssey 2020
- The 2020 Personalized Voice Trigger Challenge: Open Database, Evaluation Metrics and the Baseline Systems, arXiv 2021
- Duke Kunshan University, Post-Challenge Audio-Visual Wake Word Spotting System for the MISP 2021 Challenge, ICASSP 2023
- DiPCo, Dinner Party Corpus, 2019
- CHiME-6 Challenge: Tackling Multispeaker Speech Recognition for Unsegmented Recordings, 2020
- Cornell et al., Recent Trends in Distant Conversational Speech Recognition: A Review of CHiME-7 and 8 DASR Challenges, arXiv 2025
- Johns Hopkins University, Situated Understanding of Errors in Older Adults’ Interactions with Voice Assistants: A Month-Long, In-Home Study, arXiv 2024
- Cohn et al., speech recognition errors in older adults with and without dementia using Amazon Alexa, arXiv 2026
- U.S. District Court N.D. Cal., Kumandan v. Google, order granting final approval, October 13, 2026
- Lopez v. Apple, Lopez Voice Assistant Settlement administrator site, 2025-2026
- Google, More information about our processes to safeguard speech data, July 2019
- Apple Newsroom, Improving Siri’s privacy protections, August 2019
- Apple Newsroom, Our longstanding privacy commitment with Siri, January 2025
- Data watch: most technical measurements in this roundup date from 2018 to 2023, because smart speaker makers publish relative improvements and architecture notes rather than absolute false-accept rates for shipping devices; treat every 2018-2021 figure as most recent available data, not current product performance. The ‘up to 19 times per day’ figure that circulates for the Northeastern study is a press headline, not a rate the study reports. The accidental triggers paper gives 99 Alexa triggers in its text and 100 in a summary table row. Google’s multi-channel paper is listed with 27.2% and 31.8% false reject reductions on its Google Research page; other listings of related versions show different figures, so cite the version you read. The Johns Hopkins study gives its weekly error rate as 20-30% in the text and 21-26% in a figure caption. The Voicebot 2018 and The Manifest surveys are not comparable (the second includes phones and button presses). Google’s settlement payments had not been distributed at the time of the October 2026 order, and the order reports claims as of September 17, 2026, plus more than 40,000 late claims that may be accepted.
Last updated: October 24, 2026. We update this roundup quarterly, and the next refresh is expected when the Kumandan v. Google settlement payments are distributed and the ICASSP 2027 proceedings are released.