Bluetooth earbuds on the default SBC codec averaged 259.8 ms of audio delay in RTINGS lab tests, while the tightest osu!mania judgment window at top difficulty is only plus or minus 13.9 ms. That gap is the reason every serious rhythm game ships a calibration screen. Windows itself is no longer the bottleneck: its audio engine adds 1.3 ms in Windows 10 and later (Microsoft, Low Latency Audio 2024), and a 2024 laboratory study put the average human just noticeable difference for audio latency at 48.8 ms from a zero baseline (Schmid et al., Audio Mostly 2024). This roundup focuses on audio output delay and offset calibration; for network ping, input lag and display response, see our separate game latency statistics. We aggregated data from Microsoft, Google’s Android compatibility program, RTINGS, the osu! and StepMania source code, a Harmonix patent filing, and peer-reviewed studies in PLoS ONE, Scientific Reports, IEEE and ACM proceedings.
TL;DR
- Windows 10 and later cut audio engine latency to 1.3 ms, from about 12 ms for floating-point apps (Microsoft, Low Latency Audio 2024).
- The newest Android Media Performance Class level, MPC 37, caps tap-to-tone latency at 65 ms (Google, Android 17 MPC requirements 2026).
- Truly wireless earbuds on SBC averaged 259.8 ms of latency across five products (RTINGS, Test Bench 1.6 Latency R&D 2023).
- osu!mania’s Perfect window shrinks from plus or minus 22.4 ms to plus or minus 13.9 ms as difficulty rises (ppy, osu! source code 2026).
- StepMania’s default judgment windows run from 22.5 ms (W1) to 180 ms (W5) (StepMania 5.1 source code).
- StepMania ships with a default global audio offset of -8 ms (StepMania 5.1 source code).
- A Harmonix filing describes external video lag of about 85 ms versus about 45 ms for external audio (Harmonix, US20090310027A1 2009).
- Mean just noticeable difference for audio latency: 48.8 ms from a 0 ms base, 27.1 ms from a 64 ms base (Schmid et al., Audio Mostly 2024).
- Non-musicians needed 180 ms of key-to-sound delay to detect it, musicians 102 ms (van Vugt and Tillmann, PLoS ONE 2014).
1. Windows Audio Output Latency: Where the Desktop Delay Comes From
The headline number for PC rhythm players is that the Windows 10 audio engine adds just 1.3 ms of render latency for all applications (Microsoft, Low Latency Audio 2024). The engine stopped being the problem years ago. What still matters is the buffer: unless a game explicitly asks for smaller buffers through WASAPI (the Windows Audio Session API) or AudioGraph, it gets the 10 ms default, and the driver decides how small it can go.
That makes latency on PC a driver and game-engine question rather than an operating-system one. Microsoft also notes the trade-off: a 1 ms buffer wakes the CPU every millisecond instead of every 10 ms, which raises power draw and shortens laptop battery life (Microsoft Learn, Low Latency Audio). For how buffer sizes play out in music production software, see our DAW audio buffer size statistics.
| Metric | Value | Source |
|---|---|---|
| Windows audio engine render latency before Windows 10, apps using floating-point vs integer audio data | About 12 ms vs about 6 ms | Microsoft, Low Latency Audio documentation 2024 |
| Windows audio engine render latency in Windows 10 and later, all applications | 1.3 ms | Microsoft, Low Latency Audio documentation 2024 |
| Default buffer size Windows applications use unless they request a smaller one | About 10 ms | Microsoft, Low Latency Audio documentation 2024 |
| Round-trip latency reduction in Windows 10 compared with Windows 8.1, with no app or driver changes | 4.5 to 16 ms (4.5 ms for integer-data apps, 16 ms for floating-point apps) | Microsoft, Low Latency Audio documentation 2024 |
| Buffer range supported by the built-in Microsoft HD Audio driver | 128 samples (2.66 ms at 48 kHz) to 480 samples (10 ms at 48 kHz) | Microsoft, Low Latency Audio documentation 2024 |
Context note: a smaller buffer helps only if the game asks for it. A rhythm game that opens a standard shared-mode stream without requesting a smaller period inherits the default buffer, whatever the sound card supports.
2. Mobile Audio Latency: Android’s Guarantees and Their Ceilings
Android exposes latency as promises rather than measurements: there is no runtime API that reports actual audio latency, only feature flags that set a ceiling. The gap between those ceilings is wide. A device can advertise “low latency” audio while allowing up to 45 ms of continuous output delay, more than double the 20 ms round-trip bar for the separate pro audio flag (Android NDK, Audio latency).
The Media Performance Class (MPC) tables, which group devices by capability level, show the ceilings tightening over time: tap-to-tone latency, the delay from a screen touch to an audible sound, falls from 110 ms at the earliest levels to 65 ms at MPC 37 (Android 17 Media Performance Class). Even 65 ms is several times wider than the tightest judgment windows in section 4, which is why mobile rhythm games lean so heavily on calibration.
| Metric | Value | Source |
|---|---|---|
| Continuous output latency promised by the android.hardware.audio.low_latency feature flag | 45 ms or less | Google, Android NDK Audio Latency guide |
| Continuous round-trip latency promised by the android.hardware.audio.pro feature flag | 20 ms or less | Google, Android NDK Audio Latency guide |
| Round-trip latency that musicians generally require, according to Google | 10 ms | Google, Android NDK Audio Latency guide |
| Maximum tap-to-tone latency, MPC 1 and 10 / MPC 20 to 31 / MPC 33 to 35 | 110 / 100 / 80 ms | Google, Android 17 Media Performance Class requirements 2026 |
| Maximum round-trip audio latency, MPC 33 to 35 (first levels to set one) | 80 ms | Google, Android 17 Media Performance Class requirements 2026 |
| Maximum tap-to-tone latency and maximum round-trip latency, MPC 37 | 65 ms each | Google, Android 17 Media Performance Class requirements 2026 |
Context note: these are certification ceilings, not measurements. A phone can pass its performance class at 65 ms and still sit far outside a 20 ms judgment window, so the in-game offset has to absorb the rest.
3. Bluetooth and Wireless Headphone Delay
Bluetooth is the single largest latency source most players will ever add. RTINGS’ re-measurement of 27 headphones found that truly wireless earbuds on SBC, the default Bluetooth codec (the format that compresses audio for the wireless link), averaged 259.8 ms (RTINGS, Test Bench 1.6 Latency R&D). Gaming headsets were faster over Bluetooth, but the real difference came from their own USB wireless dongles, which measured 22 to 42 ms.
The practical reading: a calibration offset can hide a fixed Bluetooth delay, but it cannot shorten it, and any reconnection can shift the delay again. Codec choice matters more than headphone brand, as the aptX Low Latency row below shows. For how processing inside the earcup adds its own delay, see our noise-cancelling headphone DSP latency statistics.
| Metric | Value | Source |
|---|---|---|
| Average SBC latency, 5 truly wireless earbuds | 259.8 ms | RTINGS, Test Bench 1.6 Latency R&D article 2023 |
| Average SBC latency, 7 non-gaming headphones that are not earbuds | 257.6 ms | RTINGS, Test Bench 1.6 Latency R&D article 2023 |
| Average SBC latency, 6 gaming headsets connected over Bluetooth | 170.3 ms | RTINGS, Test Bench 1.6 Latency R&D article 2023 |
| Latency range of gaming headsets on their own USB wireless dongles | 22 to 42 ms | RTINGS, Test Bench 1.6 Latency R&D article 2023 |
| Plantronics BackBeat Pro 2 latency: wired / SBC / aptX / aptX Low Latency | 7 / 173 / 166 / 34 ms | RTINGS, SBC vs aptX article 2017 (updated 2025) |
| Frame durations of the LC3 codec used by Bluetooth LE Audio | 10 ms (optimized setting) or 7.5 ms (for compatibility with legacy Bluetooth Classic intervals) | Baert et al., IEEE 2024 (Experimental Validation of a Bluetooth-Based Speech Audio Broadcasting Model) |
Context note: RTINGS’ results apply to Windows under its own rig; it reports measurement precision under 1.0 ms. A shorter LC3 frame does not mean a proportionally shorter delay: in a Bluetooth SIG interview, engineer Peter Liu said the 7.5 ms mode does not result in proportionally lower latency (Bluetooth SIG, A behind-the-scenes look at LE Audio). Most recent available data for the per-category averages: RTINGS, 2023.
4. Timing Windows: How Much Error the Games Forgive
A judgment window is the span of timing error, early or late, that still earns a given grade. The open-source code of two major rhythm games shows how narrow they get. In osu!, every window is a plus-or-minus value, and the osu!mania Perfect window at overall difficulty 10 allows only 13.9 ms of error in either direction (ppy, ManiaHitWindows.cs). Overall difficulty (OD) is the per-chart setting from 0 to 10 that tightens the windows.
Put the numbers side by side and the calibration problem is obvious. The 259.8 ms SBC earbud average from section 3 is 18.7 times that 13.9 ms half-window (259.8 / 13.9 = 18.7), and it is larger than even osu!‘s loosest scoring window at the easiest setting, 200 ms (ppy, OsuHitWindows.cs). Without an offset, a player who hits exactly on the sound they hear through such earbuds would land outside every scoring window and inside the miss range.
| Metric | Value | Source |
|---|---|---|
| osu! (standard mode) Great window at overall difficulty 0 / 5 / 10 | ±80 / ±50 / ±20 ms | ppy, osu! source code (OsuHitWindows.cs) 2026 |
| osu! Ok window at overall difficulty 0 / 5 / 10 | ±140 / ±100 / ±60 ms | ppy, osu! source code (OsuHitWindows.cs) 2026 |
| osu! Meh window at overall difficulty 0 / 5 / 10 | ±200 / ±150 / ±100 ms | ppy, osu! source code (OsuHitWindows.cs) 2026 |
| osu! miss window, fixed regardless of difficulty | 400 ms | ppy, osu! source code (OsuHitWindows.cs) 2026 |
| osu!mania Perfect window at overall difficulty 0 / 5 / 10 | ±22.4 / ±19.4 / ±13.9 ms | ppy, osu! source code (ManiaHitWindows.cs) 2026 |
| osu!mania Great window at overall difficulty 0 / 5 / 10 | ±64 / ±49 / ±34 ms | ppy, osu! source code (ManiaHitWindows.cs) 2026 |
| StepMania default judgment window values, W1 (best) to W5 (worst) | 22.5 / 45 / 90 / 135 / 180 ms | StepMania 5.1 source code (Player.cpp) |
| Example input window for a rhythm-action game in a Harmonix patent filing | ±50 ms | Harmonix, US20090310027A1 2009 |
Context note: the osu!mania values are the defaults before speed and difficulty multipliers; a separate classic calculation applies with legacy scoring. StepMania’s code also sets wider defaults for held notes (250 ms) and rolls (500 ms) (StepMania, Player.cpp).
5. Calibration and Offsets: Splitting Audio Lag From Video Lag
Calibration exists because audio and video rarely arrive late by the same amount. A Harmonix patent filing for separate audio and video lag calibration describes an external video system adding about 85 ms of lag against about 45 ms for external audio, so the two would reach the player about 40 ms apart without correction (Harmonix, US20090310027A1). The filing’s own example input window is plus or minus 50 ms, so a 40 ms mismatch would eat most of it.
Calibration also has to separate what the device adds from what the player adds. Harmonix’s worked example subtracts controller and input/output lag from a measured round trip before storing an audio offset, which is why a single tap test that lumps everything together can mislead. Even professional measurement drifts: RTINGS found that skipping a connection reset between Bluetooth latency tests added about 50 ms of error, so an offset set before a reconnection may no longer be right after it.
| Metric | Value | Source |
|---|---|---|
| StepMania default global audio offset | -8 ms (-0.008 s) | StepMania 5.1 source code (PrefsManager.cpp) |
| Lag added by an external audio system, Harmonix example | About 45 ms | Harmonix, US20090310027A1 2009 |
| Lag added by an external video system, Harmonix example | About 85 ms | Harmonix, US20090310027A1 2009 |
| Video lag that DLP and rear-projection HDTVs may introduce | Up to 100 ms | Harmonix, US20090310027A1 2009 |
| Worked audio calibration: 80 ms measured round trip minus 5 ms controller lag and 7 ms input/output lag | 68 ms audio system lag | Harmonix, US20090310027A1 2009 |
Context note: the Harmonix figures are illustrative examples in a 2008 filing (published December 17, 2009, and later abandoned), not survey measurements of televisions. Most recent available data: Harmonix, 2009. The StepMania offset setting is stored in seconds (StepMania, PrefsManager.cpp).
6. Human Timing Perception Thresholds
The perception research explains why windows sit where they do. In a 2024 Audio Mostly study with 34 analyzed participants, the mean just noticeable difference (JND, the smallest change a listener reliably detects) for audio latency was 48.8 ms from a 0 ms starting point, but only 27.1 ms from a 64 ms starting point (Schmid et al., Measuring the Just Noticeable Difference for Audio Latency). People are most sensitive to latency changes in the range where real devices already sit.
Detecting that a single sound is late is much harder than tapping in time. In van Vugt and Tillmann’s study, non-musicians needed 180 ms of key-to-sound delay before reporting it (PLoS ONE), yet tappers typically land 25 to 50 ms ahead of a metronome, a habit researchers call negative mean asynchrony (Hove, Keller and Krumhansl, Perception and Psychophysics). Players may not hear 50 ms of lag as lag, but their scores will show it.
| Metric | Value | Source |
|---|---|---|
| Just noticeable difference in audio latency, starting from a 0 ms base | 48.8 ms (SD 33.0; range 10 to 129 ms) | Schmid et al., Audio Mostly 2024 |
| Just noticeable difference in audio latency, starting from a 64 ms base | 27.1 ms (SD 17.0) | Schmid et al., Audio Mostly 2024 |
| Just noticeable difference in audio latency, starting from a 512 ms base | 77.1 ms (SD 57.6) | Schmid et al., Audio Mostly 2024 |
| Delay between a key press and its sound needed for detection, non-musicians | 180 ± 104 ms | van Vugt and Tillmann, PLoS ONE 2014 |
| Delay between a key press and its sound needed for detection, musicians (pianists and brass players) | 102 ± 65 ms | van Vugt and Tillmann, PLoS ONE 2014 |
| Standard deviation of tap timing against a metronome, pianists / brass players / non-musicians | 19.9 / 19.3 / 37.4 ms | van Vugt and Tillmann, PLoS ONE 2014 |
| Typical amount by which taps precede the metronome beat (negative mean asynchrony) | 25 to 50 ms | Hove, Keller and Krumhansl, Perception and Psychophysics 2007 |
| Gap sufficient for highly trained listeners to tell which of two sounds came first | Approximately 20 ms | Hove, Keller and Krumhansl, Perception and Psychophysics 2007 |
Context note: participants in the 2024 JND study were 19 to 29 years old and none were professional musicians. Most recent available data for the keypress and tapping figures: van Vugt and Tillmann, 2014, and Hove et al., 2007.
7. Skill, Experience, and Tolerance for Added Audio Delay
Skill changes both what players notice and how they feel about it. People whose taps locked onto the beat more tightly also detected audio delay sooner, ρ(49) = -.53, p < .0001 (van Vugt and Tillmann, PLoS ONE 2014), and musical sophistication points the same way in the 2024 JND data in the table below. Trained ears are less forgiving, which is why competitive players re-calibrate after every hardware change.
The closest controlled game study is not a rhythm game: Halbhuber et al. added 0, 40, 270 and 500 ms of audio latency to a first-person shooter and found no significant score effect, but expert players reported clearly more tension at 500 ms (Halbhuber et al., Mensch und Computer 2022). Delay hurts the experience before it hurts the scoreboard. A 2025 study adds perspective on human precision itself: a trained sea lion kept a steadier beat than all 10 non-musician humans tested (Cook et al., Scientific Reports). For how timing feeds immersion in VR titles, see our VR game audio presence statistics.
| Metric | Value | Source |
|---|---|---|
| Effect of added audio latency (0, 40, 270, 500 ms) on score, 24 first-person shooter players | Not significant, F(2,66) = 1.84, p = 0.154 | Halbhuber et al., Mensch und Computer 2022 |
| Expert players’ mean score with 0 ms vs 500 ms of added audio latency | 348.7 vs 324.1 points | Halbhuber et al., Mensch und Computer 2022 |
| Expert players’ self-reported tension with 0 ms vs 500 ms of added audio latency | 0.56 vs 1.65 (p = 0.041) | Halbhuber et al., Mensch und Computer 2022 |
| Correlation between musical sophistication score (Gold-MSI) and latency JND | r(32) = -0.345, p = 0.046 | Schmid et al., Audio Mostly 2024 |
| Correlation between delay-detection threshold and synchronization accuracy | ρ(49) = -.53, p < .0001 | van Vugt and Tillmann, PLoS ONE 2014 |
| Standard deviation of inter-tap intervals after the metronome stops, pianists / brass players / non-musicians | 17.9 / 19.6 / 31.4 ms | van Vugt and Tillmann, PLoS ONE 2014 |
| Mean phase angle of 10 non-musician humans against a metronome at 112 / 120 / 128 bpm | -15.9 / -14.7 / -21.4 degrees (negative means ahead of the beat) | Cook et al., Scientific Reports 2025 |
| Interval standard deviation of a trained sea lion compared with the 10 non-musician humans | Lower than every human participant at all three tempos | Cook et al., Scientific Reports 2025 |
Context note: the 2022 latency study used a shooter, not a rhythm game, so its no-effect result on score should not be read as proof that rhythm scores are immune to delay. Most recent available data: Halbhuber et al., 2022.
Summary: Rhythm Game Audio Latency and Calibration by the Numbers
| Metric | Value | Source |
|---|---|---|
| Windows audio engine render latency, Windows 10 and later | 1.3 ms | Microsoft, Low Latency Audio documentation 2024 |
| Default Windows application buffer size | About 10 ms | Microsoft, Low Latency Audio documentation 2024 |
| Android pro audio flag round-trip ceiling | 20 ms or less | Google, Android NDK Audio Latency guide |
| Maximum tap-to-tone latency, MPC 37 | 65 ms | Google, Android 17 Media Performance Class requirements 2026 |
| Average SBC latency, truly wireless earbuds | 259.8 ms | RTINGS, Test Bench 1.6 Latency R&D article 2023 |
| Gaming headset latency on their own USB wireless dongles | 22 to 42 ms | RTINGS, Test Bench 1.6 Latency R&D article 2023 |
| LC3 codec frame durations | 10 ms or 7.5 ms | Baert et al., IEEE 2024 |
| osu!mania Perfect window at overall difficulty 10 | ±13.9 ms | ppy, osu! source code (ManiaHitWindows.cs) 2026 |
| StepMania W1 to W5 judgment windows | 22.5 to 180 ms | StepMania 5.1 source code (Player.cpp) |
| StepMania default global audio offset | -8 ms | StepMania 5.1 source code (PrefsManager.cpp) |
| External audio vs external video lag, Harmonix example | About 45 ms vs about 85 ms | Harmonix, US20090310027A1 2009 |
| Just noticeable difference in audio latency, 0 ms base | 48.8 ms | Schmid et al., Audio Mostly 2024 |
| Key-to-sound delay detection, non-musicians vs musicians | 180 ms vs 102 ms | van Vugt and Tillmann, PLoS ONE 2014 |
| Typical lead of taps ahead of the beat | 25 to 50 ms | Hove, Keller and Krumhansl, Perception and Psychophysics 2007 |
| Non-musician phase angle against a metronome at 120 bpm | -14.7 degrees | Cook et al., Scientific Reports 2025 |
Methodology and Sources
- Microsoft Learn, Low Latency Audio (Windows drivers), dated December 13, 2024, updated March 26, 2025: engine, buffer and driver latency figures.
- Google, Android NDK Audio latency guide: feature-flag ceilings and the 10 ms round-trip figure Google says musicians generally require.
- Google, Android 17 Media Performance Class requirements and the Android 17 Compatibility Definition Document: tap-to-tone and round-trip latency ceilings by performance class.
- RTINGS, Test Bench 1.6 Latency R&D article, published November 2, 2023: per-category Bluetooth and dongle averages across 27 headphones.
- RTINGS, SBC vs aptX, first published August 15, 2017, last updated August 1, 2025: single-headset codec comparison.
- Baert, Pittevils, Moons and Hoebeke (imec-IDLab-UGent and Cochlear Technology Centre), Experimental Validation of a Bluetooth-Based Speech Audio Broadcasting Model, IEEE 2024: LC3 frame durations.
- Bluetooth SIG, A behind-the-scenes look at Bluetooth LE Audio, September 28, 2020: context on the 7.5 ms LC3 mode.
- ppy, osu! source code: OsuHitWindows.cs and ManiaHitWindows.cs, read November 2026; windows are plus-or-minus values per the shared HitWindows class.
- StepMania 5.1 source code: Player.cpp (timing windows) and PrefsManager.cpp (global offset).
- Harmonix Music Systems, US20090310027A1, Systems and methods for separate audio and video lag calibration in a video game, filed June 16, 2008, published December 17, 2009.
- Schmid, Ambros, Bogon and Wimmer, Measuring the Just Noticeable Difference for Audio Latency, Audio Mostly 2024 (ACM).
- van Vugt and Tillmann, Thresholds of Auditory-Motor Coupling Measured with a Simple Task in Musicians and Non-Musicians, PLoS ONE, February 3, 2014.
- Hove, Keller and Krumhansl, sensorimotor synchronization study, Perception and Psychophysics 69(5), 2007.
- Halbhuber, Köhler, Schmidbauer, Wiese and Henze, The Effects of Auditory Latency on Experienced First-Person Shooter Players, Mensch und Computer 2022 (ACM).
- Cook, Hood, Rouse and Reichmuth, Sensorimotor synchronization to rhythm in an experienced sea lion rivals that of humans, Scientific Reports, 2025.
- Data watch: several themes rest on older data, flagged in each section (Harmonix 2009, Hove et al. 2007, van Vugt and Tillmann 2014, Halbhuber et al. 2022, RTINGS 2017 and 2023). RTINGS measures on a Windows rig, so phone-side Bluetooth stacks may differ, and it excluded AAC after an unreliable macOS reading (about 45 ms on AirPods Pro versus about 160 ms on SBC under Windows 11). Halbhuber et al. report the score ANOVA as F(2,66) although the design had four latency levels, and their introduction gives 400 ms where methods and results use 500 ms. Windows and Android figures are ceilings and defaults, not field measurements of real devices; no public dataset of rhythm-game players’ actual calibration offsets was found, so offset figures come from defaults and patent examples. The MPC tables may gain a new level with the next Android release.
- Last updated: November 6, 2026. We update this roundup quarterly, and the next refresh is expected when Google publishes the Media Performance Class requirements for the next Android release.