Kinect sold 35 million units and its manufacturing still ended in 2017 (MCV, Microsoft officially ceases Kinect production 2017), the clearest proof that a microphone in the box does not make voice a standard game control. Steam now returns 190 results for its Voice Control tag (Steam store search, October 31, 2026), and more than 95 million PS5 consoles have been sold in to retailers for a platform whose DualSense controller has a built-in microphone (Sony Interactive Entertainment, Business Data and Sales 2026). Recognition remains uneven: second-language English speakers face error gaps of 2.6 to 5.9 times across current models (Roll and Graham, SLaTE 2025), a pattern we also track in our speech recognition accent bias statistics. We aggregated data from University of Melbourne game research, Disney Research, Sony, Nintendo, and peer-reviewed speech science papers to separate what players actually do with their voices from what platform holders claimed.
TL;DR
- Kinect sold 35 million units before manufacturing ended in 2017 (MCV 2017).
- Microsoft claimed an average of 120 voice commands a month in 2014 (Shacknews 2014).
- Phasmophobia, which lets players speak to its ghosts, passed 25 million sales across all platforms (Kinetic Games via WorthPlaying 2025).
- Second-language English speakers saw error gaps of 2.6 to 5.9 times across 10 recognition models (Roll and Graham, SLaTE 2025).
- Accented speech scored a 30-50% word error rate on Wav2Vec 2.0, against 2-8% for native speakers (Bandodkar et al., AAAI 2024).
- A full spoken exchange with an AI game character averaged 6.909 seconds, with a worst case of 24.362 seconds (Korkiakoski et al. 2025).
- English dictation ran 2.93 times faster than typing on a phone keyboard, 153 vs 52 words per minute (Ruan et al. 2017).
- A keyword recognizer built for children reached 85% accuracy, and a commercial recognizer was 35% less accurate (Disney Research 2016).
- PS5 Voice Command remains a preview limited to two countries and one language (PlayStation Support 2026).
1. Forty Years of Rising and Falling Interest
106 new voice interaction videogames shipped in 2017 alone (Allison et al., Design Patterns for Voice Interaction in Games, CHI PLAY 2018), which sounds like a breakthrough until you read the history behind it. The same research group found that voice has “never achieved sustained, widespread success” and instead moved through seven phases tied to new platforms and enabling technologies (Allison, Carter and Gibbs, Word Play, Games and Culture 2020).
Karaoke is the instructive peak. It asked players to perform with their voice, not to be understood word by word. The audiogame branch on smart speakers follows a related logic: designers mostly reduced speech to choosing from a short list, which keeps the recognition problem small.
| Metric | Value | Source |
|---|---|---|
| Games in the largest published voice game survey, 1973 to 2018 | 449 videogames plus 22 audiogames | Allison et al., Design Patterns for Voice Interaction in Games, CHI PLAY 2018 |
| New voice interaction videogames released in one year | 106 in 2017 | Allison et al., CHI PLAY 2018 |
| Karaoke games released from 2003 to 2012, counting multi-platform releases once | More than 120, of which more than 70 came in 2008 to 2010 | Allison et al., Word Play, Games and Culture 2020 |
| Games that used the Nintendo 64 Voice Recognition Unit, 1998 to 1999 | 2 | Allison et al., Word Play 2020 |
| Major releases using Kinect voice recognition at the time of the study | More than 30 | Allison et al., Word Play 2020 |
| Alexa audiogames whose core interaction was choosing from spoken options | 17 of 22 surveyed | Allison et al., CHI PLAY 2018 |
Most recent available data: Allison et al., 2018 and 2020. We found no newer census of voice games, so these counts describe the field up to early 2018.
2. Microphones Everywhere, Voice Commands Nowhere
More than 95 million PS5 consoles have been sold in to retailers (Sony Interactive Entertainment, Business Data and Sales, June 30, 2026), and the DualSense controller has an internal microphone. That installed base has not produced a voice-first platform. PlayStation Support still describes Voice Command as a preview for accounts in the United States and United Kingdom, in English only, and aimed at system tasks such as opening games and capturing clips rather than at gameplay.
The contrast with the home is the useful number here. 57% of voice command users use voice commands daily (NPR and Edison Research, Smart Audio Report 2022). People are comfortable speaking to devices, so the barrier in games is not the habit of talking to machines. Buttons are silent, private and immediate, and voice has to beat that.
| Metric | Value | Source |
|---|---|---|
| Kinect units sold in the first 60 days on the market | 8 million | MCV, Microsoft officially ceases Kinect production 2017 |
| Kinect lifetime unit sales | 35 million | MCV 2017 |
| Share of Kinect owners using it, as stated by Microsoft (company claim, not independently verified) | 80% | Shacknews reporting Yusuf Mehdi of Microsoft, May 16, 2014 |
| Average voice commands per month, as stated by Microsoft (company claim) | 120 | Shacknews reporting Microsoft, 2014 |
| Total voice commands issued, as stated by Microsoft (company claim) | More than 1 billion | Shacknews reporting Microsoft, 2014 |
| PS5 units sold in to retailers, cumulative to June 30, 2026 | More than 95 million | Sony Interactive Entertainment, Business Data and Sales 2026 |
| Nintendo Switch 2 units sold, cumulative to June 30, 2026 (built-in microphone for GameChat) | 23.68 million | Nintendo, Consolidated Hardware Sales 2026 |
| Americans aged 18+ who use a voice assistant on any device | 62% | NPR and Edison Research, Smart Audio Report 2022 |
Most recent available data: Shacknews, 2014 for Kinect usage, and the Smart Audio Report, 2022 for voice assistants. We found no Sony or Nintendo figure for how many players use a microphone for game commands, and MCV reports the only Kinect sales trail we could verify. The Switch 2 microphone also filters out game sound and adapts to the player’s distance from the system (Nintendo, GameChat features), and its voice chat tops out at 12 participants.
3. Today’s Voice Game Catalogue
25 of the first 30 results on Steam’s Voice Control tag were dated 2025 or 2026 (our tally of the Steam store search, October 31, 2026), 15 of them in 2026 alone. That is one page of a store listing, not a release census, so it does not prove a boom on its own. It does show that the microphone is still being used as a design hook by new titles.
Phasmophobia is the scale case. It launched on consoles without recognition, then developer Kinetic Games added a console voice update for PlayStation 5, PS VR2 and Xbox Series X|S after several months, using a recognizer called VOSK. A studio doing that work for a shipped game treats voice as part of the product, not as a bonus.
| Metric | Value | Source |
|---|---|---|
| Steam store results for the Voice Control tag, including demos and utility software | 190 | Steam store search, October 31, 2026 |
| Steam titles counted under the same tag by a third-party tracker (undated snapshot) | 113 | DataHumble, Voice Control Games 2026 |
| Phasmophobia sales across all platforms | More than 25 million | Kinetic Games via WorthPlaying, July 1, 2025 |
| Phasmophobia sales on consoles | More than 3 million | Kinetic Games via WorthPlaying, April 23, 2025 |
| Languages supported by Phasmophobia console voice recognition at launch | 7 | Kinetic Games, Console Voice Recognition patch notes 2025 |
Phasmophobia sales cover every buyer, not every player who spoke to a ghost. We found no source reporting what share of its players enable the microphone.
4. Recognition Accuracy Is Not Shared Equally
Second-language English speakers faced error gaps of 2.6 to 5.9 times compared with first-language speakers across 10 models and more than 300 English varieties (Roll and Graham, Scaling Conformation Bias in Automatic Speech Recognition, SLaTE 2025). Word error rate (WER) is the share of spoken words a recognizer gets wrong, so a lower number is better. The paper’s uncomfortable finding is that overall improvement does not close the gap, since relative disparity grew as models improved for first-language speakers.
For a game shipped worldwide, this turns a technical metric into a question of who can play. Radio General’s own store page says it works only in English and works best with a North American accent. Context is one proposed remedy: a Sony patent filed on May 28, 2022 and granted on May 13, 2025 uses the current game state to help decide what the player said.
| Metric | Value | Source |
|---|---|---|
| Mean WER of the best overall model tested, Whisper-large | 6.9% | Roll and Graham, SLaTE 2025 |
| Mean WER of the smallest Whisper model, Whisper-tiny | 13.6% | Roll and Graham, SLaTE 2025 |
| WER gap between second-language and first-language English speakers across models | 2.6 to 5.9 times | Roll and Graham, SLaTE 2025 |
| WER of speakers of tonal first languages vs speakers of Indo-European first languages | 2 to 3 times higher | Roll and Graham, SLaTE 2025 |
| Whisper-large WER in one corpus (Speech Accent Archive), first-language vs second-language speakers | 0.8% vs 3.7% | Roll and Graham, SLaTE 2025 |
| Wav2Vec 2.0 WER, accented speech vs native speakers | 30-50% vs 2-8% | Bandodkar et al., AAAI 2024 |
| Relative improvement in macro-averaged WER from fairness-focused retraining vs the pretrained Whisper model (26 accent groups analyzed) | 58.7% | Swain et al., arXiv 2025 |
None of these tests used recordings from gameplay. Game audio adds excitement, shouting and leakage from speakers, so treat these error rates as a lower bound on difficulty, not a forecast.
5. Speech Is Fast, Recognition Is Not
Speech beat the phone keyboard by 2.93 times, at 153 vs 52 words per minute (Ruan et al., 2017). That advantage explains why voice keeps returning to games, but dictation and game control are different tasks. A dictated sentence can arrive a second later without harm, while a dodge cannot. Most recent available data: Ruan et al., Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 2017.
Large language models widen the gap. In a VR test with 18 participants, the wait for an AI character’s reply was dominated by the model and voice output, not the listening stage. That puts free conversation well outside the latency budgets of action games, and it is why voice-driven characters so far suit slower genres. Our AI in gaming statistics cover the wider industry picture, while the figures here concern only the player’s spoken input.
| Metric | Value | Source |
|---|---|---|
| English speech vs phone keyboard text entry speed | 153 vs 52 words per minute | Ruan et al., 2017 |
| Corrected error rate, speech vs phone keyboard | 5.30% vs 11.22% | Ruan et al., 2017 |
| Uncorrected error rate, speech vs phone keyboard | 1.30% vs 0.79% | Ruan et al., 2017 |
| Typical spoken length of a one-word game command (“go” or “jump”) | About 300 milliseconds | Disney Research via Phys.org, 2016 |
| Speech-to-text stage in an AI character test, mean vs worst case | 1.365 vs 3.475 seconds | Korkiakoski et al., arXiv 2025 |
| Language model stage (GPT-4 Turbo) in the same test, mean vs worst case | 3.113 vs 14.918 seconds | Korkiakoski et al., arXiv 2025 |
| Text-to-speech stage in the same test, mean vs worst case | 2.376 vs 12.404 seconds | Korkiakoski et al., arXiv 2025 |
| Full exchange, from player speech to character reply, mean vs worst case | 6.909 vs 24.362 seconds | Korkiakoski et al., arXiv 2025 |
A hybrid design narrows the job the model must do. One 2025 study had players paraphrase predefined dialogue options aloud while a language model matched them, and its 14 participants reported more enjoyment and perceived freedom than with point-and-click (Wevelsiep et al., IWSDS 2025). Its authors still list latency as a priority for future work.
6. Children, Speech Therapy and Access
A keyword recognizer tuned for children reached 85% accuracy, while a commercial recognizer was 35% less accurate in a two-player game called Mole Madness (Disney Research, 2016, reported by Phys.org). The lesson is that narrowing the vocabulary to two words did more than a general-purpose engine could. This matches the Game Accessibility Guidelines, which recommend recognition based on a small vocabulary, and at the advanced level on simply hitting a volume threshold.
Speech therapy games show the limit plainly. Their recognizers judge the very speech that is hard to recognize, and in one trial the machine agreed with humans on fewer than half of the correct productions from children with childhood apraxia of speech, a motor speech disorder (Ahmed et al., International Journal of Speech-Language Pathology 2018). The same Guidelines list “speech input is not required” as a basic rule, and Xbox Accessibility Guideline 107 asks that gameplay-critical speech have a digital alternative. Voice can open a game to some players and close it to others, which our game accessibility features statistics explore across every input type.
| Metric | Value | Source |
|---|---|---|
| Keyword recognition accuracy in Mole Madness | 85% | Disney Research via Phys.org, 2016 |
| How much less accurate a commercial recognizer was than the keyword spotter | 35% | Disney Research via Phys.org, 2016 |
| Recognizer-human agreement on incorrect productions, typically developing children vs children with childhood apraxia of speech | 77% vs 75% | Ahmed et al., 2018 |
| Recognizer-human agreement on correct productions, by speaker: typically developing children vs children with apraxia vs speech-language pathologists | 51% vs 47% vs 71% | Ahmed et al., 2018 |
| PocketSphinx accuracy on child speech therapy audio, default vs adapted with maximum a posteriori (MAP) training | 59.4% vs 80.0% | Hair et al., ASSETS 2019 |
| Mean accuracy of a custom template-matching recognizer in the same evaluation | 75.8% | Hair et al., ASSETS 2019 |
| Primary school pupils who tested a Dutch speech-based reading game prototype | More than 40 | Radboud University, 2023 |
| Studies identified vs included in a 2026 review of 3D audio games for visually impaired players | 975 vs 56 | Kaka and Juhana, JoMMiT 2026 |
Most recent available data: Disney Research, 2016; Ahmed et al., 2018; Hair et al., 2019. These are clinical or research settings with small samples, and they report system performance only.
7. How Players Actually Respond
Phasmophobia holds 94% positive across 360,726 English reviews, while Suck Up! sits at 61% (Steam store pages, October 31, 2026). Both ask you to talk to the game, but they ask differently. In one, voice is a single feature inside a wider ghost hunt. In the other, talking your way past AI characters is the core loop, and the mixed result repeats in DON’T SCREAM, where being too loud ends the run.
Reviews cannot isolate voice from everything else in a game, so read this as a pattern, not a measurement. The academic work points the same way. A CHI PLAY 2015 study found that understanding the player’s identity in the game matters alongside recognition accuracy. A small 2018 study went further, with 5 participants showing higher engagement but lower happiness with voice than with keyboard (p = 2.005e-14 and p = 1e-3, paired t-test, Mattey and Yuksel, OverVoice). With a sample that small, treat it as a hypothesis.
| Metric | Value | Source |
|---|---|---|
| Items analyzed on player reception of voice games (166 + 2,951 + 84 + 69) | 3,270 reviews, threads and videos | Carter et al., CHI PLAY 2015 |
| Phasmophobia positive share of English Steam reviews | 94% of 360,726 | Steam store page, 2026 |
| Mage Arena positive share of English Steam reviews (spells cast by voice) | 90% of 11,561 | Steam store page, 2026 |
| Tom Clancy’s EndWar positive share of Steam reviews (2009 strategy game with voice orders) | 80% of 1,035 | Steam store page, 2026 |
| DON’T SCREAM positive share of English Steam reviews | 68% of 1,795 | Steam store page, 2026 |
| Suck Up! positive share of Steam reviews (players talk to AI characters) | 61% of 226 | Steam store page, 2026 |
| System Usability Scale score for spoken AI characters in VR (0 to 100 scale) | 79.44 | Korkiakoski et al., arXiv 2025 |
| Believability of spoken AI characters in VR (out of 10) | 6.67 | Korkiakoski et al., arXiv 2025 |
Most recent available data on player attitudes: Carter et al., 2015. We found no large player survey on voice command preference, so no share of players who prefer voice control appears in this roundup.
Summary: Voice Command and Speech Recognition in Video Games by the Numbers
| Metric | Value | Source |
|---|---|---|
| New voice interaction videogames released in one year | 106 in 2017 | Allison et al., CHI PLAY 2018 |
| Kinect lifetime unit sales | 35 million | MCV 2017 |
| Average voice commands per month, as stated by Microsoft (company claim) | 120 | Shacknews 2014 |
| PS5 units sold in to retailers, cumulative to June 30, 2026 | More than 95 million | Sony Interactive Entertainment 2026 |
| Nintendo Switch 2 units sold, cumulative to June 30, 2026 | 23.68 million | Nintendo 2026 |
| Americans aged 18+ who use a voice assistant on any device | 62% | NPR and Edison Research 2022 |
| Steam store results for the Voice Control tag | 190 | Steam store search, October 31, 2026 |
| Phasmophobia sales across all platforms | More than 25 million | Kinetic Games via WorthPlaying 2025 |
| Mean WER of the best overall model tested | 6.9% | Roll and Graham, SLaTE 2025 |
| WER gap between second-language and first-language English speakers | 2.6 to 5.9 times | Roll and Graham, SLaTE 2025 |
| Wav2Vec 2.0 WER, accented speech vs native speakers | 30-50% vs 2-8% | Bandodkar et al., AAAI 2024 |
| English speech vs phone keyboard text entry speed | 153 vs 52 words per minute | Ruan et al., 2017 |
| Full spoken exchange with an AI character, mean vs worst case | 6.909 vs 24.362 seconds | Korkiakoski et al., arXiv 2025 |
| Keyword recognition accuracy in Mole Madness | 85% | Disney Research via Phys.org, 2016 |
| Recognizer-human agreement on correct productions, children with childhood apraxia of speech | 47% | Ahmed et al., 2018 |
Methodology and Sources
We read every figure in the sources below during this update and traced company figures to the company that published them. Press reports are named where the original statement was not available to read directly.
- Allison, Carter, Gibbs and Smith, Design Patterns for Voice Interaction in Games, CHI PLAY 2018
- Allison, Carter and Gibbs, Word Play: A History of Voice Interaction in Digital Games, Games and Culture 2020
- Allison, Carter and Gibbs, A History of Voice Interaction in Games, DiGRA and FDG 2016
- Carter, Allison, Downs and Gibbs, Player Identity Dissonance and Voice Interaction in Games, CHI PLAY 2015
- MCV, Microsoft officially ceases Kinect production, October 26, 2017
- Shacknews, Including Kinect was ‘the right call,’ says Microsoft exec, May 16, 2014
- Sony Interactive Entertainment, Business Data and Sales, as of June 30, 2026
- PlayStation Support, Voice Command (Preview) on PS5
- Sony Interactive Entertainment, US patent 12300221, granted May 13, 2025
- Nintendo, Consolidated Hardware and Software Sales Units, as of June 30, 2026
- Nintendo, GameChat features for Nintendo Switch 2
- NPR and Edison Research, Smart Audio Report 2022
- Steam store search, Voice Control tag, October 31, 2026
- DataHumble, Voice Control Games
- Kinetic Games, Console Voice Recognition, Phasmophobia v0.11.1.2
- Rely on Horror, Phasmophobia on PS5 and Xbox now supports voice recognition, February 28, 2025
- WorthPlaying, Phasmophobia surpasses 25 million sales, July 1, 2025
- WorthPlaying, Phasmophobia surpasses 3 million sales on consoles, April 23, 2025
- Roll and Graham, Scaling Conformation Bias in Automatic Speech Recognition, SLaTE 2025
- Bandodkar et al., “Allot?” Is “a Lot!”, AAAI 2024
- Swain et al., Towards Fair ASR For Second Language Speakers Using Fairness Prompted Finetuning, arXiv 2025
- Ruan et al., Comparing Speech and Keyboard Text Entry for Short Messages in Two Languages on Touchscreen Phones, 2017
- Korkiakoski et al., An Empirical Evaluation of AI-Powered Non-Player Characters’ Perceived Realism and Performance in Virtual Reality Environments, arXiv 2025
- Wevelsiep et al., A Voice-Controlled Dialogue System for NPC Interaction using Large Language Models, IWSDS 2025
- Disney Research (Lehman, Wolfe, Pereira and Hodgins), as reported by Phys.org, 2016
- Ahmed et al., Speech-driven mobile games for speech therapy, International Journal of Speech-Language Pathology 2018
- Hair et al., Evaluating Automatic Speech Recognition for Child Speech Therapy Applications, ASSETS 2019, listed by Texas A&M PSI Lab
- Radboud University, Playfully practicing speaking and reading, June 28, 2023
- Kaka and Juhana, 3D Audio Games for Visually Impaired Players: A Systematic Literature Review, JoMMiT 2026
- Game Accessibility Guidelines, full list
- Microsoft, Xbox Accessibility Guideline 107
- Mattey and Yuksel, OverVoice: Voice Controlled Games Engagement Study, University of San Francisco, 2018
- Steam store pages, read October 31, 2026: Phasmophobia, Mage Arena, Tom Clancy’s EndWar, DON’T SCREAM, Suck Up!, Radio General
- Data watch: We found no platform holder figure for the share of players who use voice commands, and the Kinect figures of 80%, 120 commands and 1 billion commands are unaudited Microsoft claims from 2014. The Steam result count of 190 includes demos and utility software and rests on user-applied tags, which is one reason it differs from DataHumble’s undated count of 113. Allison et al. described six phases in 2016 and seven in 2020, and we use the later count. The AAAI 2024 native-speaker range of 2-8% is cited by its authors from earlier work, not measured in that paper, and the accent studies did not use game audio. The 2.6 to 5.9 times gap is the authors’ stated range, and average WER across corpora in that paper ran from 2.3% to 26.9%, so a single headline rate depends heavily on the recordings used. Steam review counts change daily, and some pages do not state whether a count covers English reviews only. Voice chat moderation is not covered because the Call of Duty progress reports could not be reached during this update. Sony and Nintendo refresh their hardware figures with each quarterly report.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when Nintendo publishes its next Consolidated Hardware and Software Sales Units update.