On a 96 kHz ASIO interface, measured round-trip latency climbed from 2.30 ms at a 64-sample buffer to 12.30 ms at 1,024 samples (MathWorks, Measure Audio Latency example), which puts the largest setting past the under-10 ms target long recommended for digital musical instruments (Wessel and Wright 2002). Human perception is messier than that guideline: the mean just-noticeable added latency was 48.8 ms in a 2024 key-press study (Schmid et al., Audio Mostly 2024), yet clapping duos already started to lose tempo above an 11.5 ms one-way delay (Chafe and Gurevich, AES 117th Convention 2004). This roundup covers latency and performance rather than sales, so for revenue and user counts see our DAW market statistics. We aggregated data from Microsoft Learn, the Android Compatibility Definition Document, MathWorks, Scan Pro Audio, the Audio Engineering Society, NIME, Music Perception, PLoS One, PNAS, and a dozen other primary sources.
TL;DR
- Round-trip latency measured 2.30 ms at a 64-sample buffer and 12.30 ms at 1,024 samples on a 96 kHz ASIO interface (MathWorks, Measure Audio Latency example).
- The mean just-noticeable added latency was 48.8 ms from a 0 ms baseline but only 27.1 ms on top of a 64 ms baseline (Schmid et al., Audio Mostly 2024).
- Percussionists and amateurs rated 0 ms above 20 ms and above a jittery 10 ms, but not above a steady 10 ms (Jack et al., Music Perception 2018).
- 200 ms of delayed auditory feedback slowed speech from 5.2 to 4.3 syllables per second (Ozker and Hagoort, PLoS One 2025).
- Clapping duos held tempo best at 11.5 ms one-way delay, and 85% of performances slowed at 14 ms and above (Chafe and Gurevich, AES 2004).
- Piano duos sped up at 10 ms, held steady at 20 ms and slowed at 40 ms of added latency (Washburn et al., Frontiers in Psychology 2021).
- A Ryzen 9 9950X ran 2,140 virtual-instrument voices at a 64-sample buffer versus 8,580 at 512, but plugin instances only rose from 482 to 618 (Scan Pro Audio, DAWBench).
- Windows 10 cut audio engine render latency to 1.3 ms, from about 12 ms for floating-point apps (Microsoft Learn, Low Latency Audio).
- Android pro audio devices must reach 25 ms round trip or less, 20 ms on an analog path (Android 16 Compatibility Definition Document).
- A Bluetooth LE controller link averaged 139 ms from action to sound, versus about 0.6 ms for the Bela embedded platform (McPherson et al., NIME 2016).
- In-ear monitor latency of 5 ms and below is recommended to avoid compromising performance (Shure Insights 2016, vendor guidance).
1. Latency Perception: How Much Delay Musicians Notice
There is no single magic number. In the Regensburg just-noticeable-difference study, the mean just-noticeable added latency was 48.8 ms from a zero baseline but only 27.1 ms on top of a 64 ms baseline (Schmid et al., Audio Mostly 2024). A just-noticeable difference (JND) is the smallest added delay a listener reliably detects, not the delay they are comfortable playing with, and the result says a system that already lags makes further lag easier to spot, not harder.
Playing tests set the bar lower than detection tests. In Jack et al. (Music Perception 2018), professional percussionists and amateur musicians could not rate a steady 10 ms as worse than zero, but rated 20 ms and a 10 ms delay that wobbled by plus or minus 3 ms as lower quality. For buffer settings, the takeaway is that consistency matters about as much as the raw figure: a stable small buffer beats one that occasionally stretches.
| Metric | Value | Source |
|---|---|---|
| Mean just-noticeable added latency in a key-press task, starting from 0 ms base latency | 48.8 ms (SD 33.0; individual range 10 to 129 ms) | Schmid et al., Audio Mostly 2024 |
| Mean just-noticeable added latency, starting from 64 ms base latency | 27.1 ms (SD 17.0) | Schmid et al., Audio Mostly 2024 |
| Mean just-noticeable added latency, starting from 512 ms base latency | 77.1 ms (SD 57.6) | Schmid et al., Audio Mostly 2024 |
| Correlation between musical sophistication (Gold-MSI score) and just-noticeable latency | r(32) = -0.345, p = 0.046 | Schmid et al., Audio Mostly 2024 |
| Quality rating of 0 ms compared with 10 ms, 10 ms with plus or minus 3 ms jitter, and 20 ms | Higher than jittery 10 ms and 20 ms; not higher than steady 10 ms | Jack et al., Music Perception 2018 |
| Latency target recommended for digital musical instruments | Under 10 ms | Wessel and Wright 2002, cited in McPherson et al., NIME 2016 |
| Acceptable latency for live performers, depending on instrument and monitoring setup | 1.4 ms to 42 ms | Lester and Boley, AES Convention Paper 7198, 2007 |
Context note: the negative correlation means more musically trained participants detected smaller delays. Most recent available data for the performer tolerance range and the design target: Lester and Boley (AES, 2007) and Wessel and Wright (2002). No newer peer-reviewed replication of either was found.
2. Hearing Your Own Voice Late: Speech and Vocal Monitoring
Voice is the least forgiving signal to delay, because a speaker also hears it instantly through bone conduction. Research on delayed auditory feedback (DAF), meaning hearing your own voice played back late through headphones, uses 200 ms as the disruptive benchmark: 40 speakers slowed from 5.2 to 4.3 syllables per second under a 200 ms delay (Ozker and Hagoort, PLoS One 2025). That is far beyond any sensible DAW buffer, but it exposes the mechanism: the brain uses its own audio as a timing reference, and a late copy fights the real one.
Susceptibility varies widely between people. Only 3 of 40 participants were highly susceptible to both delayed and pitch-shifted feedback, and a 2026 PNAS study linked resistance to the structure of the right arcuate fasciculus, a language pathway in the brain. For anyone monitoring a processed voice in real time, whether a vocal chain in a DAW or real-time voice effects for streaming, that variability explains why one performer shrugs off a delay that throws another. In-ear monitors are the usual fix for singers, and they bring their own exposure trade-offs, covered in our musician hearing loss statistics.
| Metric | Value | Source |
|---|---|---|
| Speech rate with immediate feedback vs with 200 ms delayed feedback (40 speakers reading sentences) | 5.2 vs 4.3 syllables per second | Ozker and Hagoort, PLoS One 2025 |
| Voice loudness with immediate feedback vs with 200 ms delayed feedback | 58.2 vs 59.7 dB | Ozker and Hagoort, PLoS One 2025 |
| Speakers highly susceptible to delayed feedback, to pitch-shifted feedback, and to both | 10 of 40, 10 of 40, and 3 of 40 | Ozker and Hagoort, PLoS One 2025 |
| Mean spoken word duration with immediate feedback vs with 200 ms delayed feedback (30 speakers) | 0.85 s vs 0.95 s (t = 7.859, P < 0.001) | Ozker et al., PNAS 2026 |
| Dysfluencies at 200 ms delay compared with 0, 25 and 50 ms delays (17 adults) | Two to three times more | Stuart et al., JASA 2002 |
| In-ear monitor latency recommended to avoid compromising performance | 5 ms and below | Shure Insights 2016 (vendor guidance) |
| Stage monitor latency beyond which delay can start to become noticeable | 10 ms | Shure Insights 2016 (vendor guidance) |
Context note: Stuart et al. (JASA 2002) is the most recent available data comparing several short delays (25 and 50 ms) against 200 ms; the 2025 and 2026 studies tested 200 ms only. The Shure figures are a manufacturer’s guidance, not a controlled study, and date from 2016.
3. Playing Together: Ensemble Timing Under Delay
Ensemble data turns latency from an annoyance into a measurable tempo drift. Below an 11.5 ms one-way delay, 74% of clapping performances sped up; at 14 ms and above, 85% slowed down (Chafe and Gurevich, AES 117th Convention 2004). A small delay works like a brake that steadies the tempo, and a large one becomes a drag the players cannot shake off.
The 2021 piano duet study points the same way with real instruments: pairs sped up at 10 ms, held steady at 20 ms and slowed progressively at 40 ms (Washburn et al., Frontiers in Psychology 2021). Its review of prior work places stable synchrony at 10 to 25 ms and general deterioration beyond 60 ms. Monitoring through a DAW adds buffer latency on top of any acoustic distance in the room, so these ranges matter for tracking sessions, not only for internet jams.
| Metric | Value | Source |
|---|---|---|
| One-way delay range tested on clapping duos (15 duos analyzed, 173 trials) | 0 to 77 ms | Chafe and Gurevich, AES 117th Convention 2004 |
| One-way delay at which mean tempo neither sped up nor slowed | 11.5 ms | Chafe and Gurevich, AES 117th Convention 2004 |
| Performances that sped up with delay below 11.5 ms vs performances that slowed at 14 ms and above | 74% vs 85% | Chafe and Gurevich, AES 117th Convention 2004 |
| Piano duo tempo at 10, 20 and 40 ms of added latency (12 pairs) | Sped up at 10 ms, steady at 20 ms, slowed at 40 ms | Washburn et al., Frontiers in Psychology 2021 |
| Share of variation in collective tempo explained by added latency | R² = 0.88 | Washburn et al., Frontiers in Psychology 2021 |
| Latency ranges from prior research: synchronous playing, tempo decreases, general deterioration | 10 to 25 ms, 25 to 60 ms, beyond 60 ms | Washburn et al., Frontiers in Psychology 2021 (review of prior studies) |
| Typical internet latency range cited for networked music performance | 20 to 100 ms | Washburn et al. 2021, citing Cáceres et al. 2008 and Cáceres and Chafe 2010 |
Context note: Chafe and Gurevich (2004) is the most recent available data for the 11.5 ms sweet spot; the 2021 duet study is consistent with it but used different delay steps.
4. Buffer Size and Round-Trip Latency
Buffer size sets the floor, not the whole bill. In the MathWorks loopback measurement on an ASIO MADIface USB interface at 96 kHz, round-trip latency rose from 2.30 ms at 64 samples to 12.30 ms at 1,024 samples. A 16-fold larger buffer produced about 5.3 times the latency (12.30 / 2.30 = 5.35), because part of the delay does not depend on the buffer at all.
Our calculation from the same table isolates that fixed part. The 64-sample measurement was 221 samples, which leaves 157 samples once the 64-sample buffer is subtracted; the 1,024-sample measurement (12.3021 ms x 96 = 1,181 samples) also leaves 157 samples. That is about 1.64 ms (157 / 96,000) of converter, driver and transfer overhead that no buffer setting removes, so each doubling simply adds the duration of the extra samples: 64 more samples at 96 kHz is 0.67 ms, exactly the gap between the 64 and 128-sample rows.
| Metric | Value | Source |
|---|---|---|
| Measured round-trip latency, 64-sample buffer, 96 kHz ASIO interface | 2.30 ms (221 samples) | MathWorks, Measure Audio Latency example |
| Measured round-trip latency, 128-sample and 256-sample buffers | 2.97 ms and 4.30 ms | MathWorks, Measure Audio Latency example |
| Measured round-trip latency, 512-sample buffer | 6.97 ms | MathWorks, Measure Audio Latency example |
| Measured round-trip latency, 1,024-sample buffer | 12.30 ms | MathWorks, Measure Audio Latency example |
| Buffer range supported by the Windows inbox HDAudio driver | 128 samples (2.66 ms at 48 kHz) to 480 samples (10 ms at 48 kHz) | Microsoft Learn, Low Latency Audio |
| Typical buffer sizes reported by Android devices | 96, 128, 160, 192, 240, 256 or 512 frames | Android Developers, Audio Latency guide |
| Typical error in the latency an interface reports, corrected by Ableton Live’s Driver Error Compensation | Around plus or minus 1 ms to 20 ms | Ableton Help Center, Driver Error Compensation FAQ |
Context note: the Ableton Driver Error Compensation FAQ warns that settings above 20 ms may cause issues and that the value must be recalculated whenever buffer size or sample rate changes. In plain words, some interfaces misreport their own delay, so recordings land late unless the DAW is told the true figure.
5. Buffer Size and CPU Headroom
The cost of a small buffer shows up in how much work the CPU can finish before each audio deadline. On Scan Pro Audio’s DAWBench virtual-instrument test, run in Reaper with an RME Babyface interface, an AMD Ryzen 9 9950X sustained 2,140 voices of polyphony at a 64-sample buffer and 8,580 at 512, about 4.0 times as many (8,580 / 2,140 = 4.01). The plugin-instance test scaled far less, from 482 to 618 instances, a 28% gain (618 / 482 = 1.28).
That split is the practical lesson: large sample-library instruments are where small buffers hurt, while effect-heavy mixing chains care less. The two CPUs also trade places as the buffer grows, as the Ryzen 9 9950X and Core Ultra 7 270K Plus result tables show: the Intel chip led at 64 samples (512 vs 482 plugin instances) and the two tied at 618 at 512 samples. Apple’s Logic Pro guide gives the matching workflow advice: record at the lowest buffer setting and raise it to the next step only if system overload alerts appear.
| Metric | Value | Source |
|---|---|---|
| DAWBench DSP plugin instances, AMD Ryzen 9 9950X, at 64 / 128 / 256 / 512-sample ASIO buffers | 482 / 562 / 594 / 618 | Scan Pro Audio, 3XS Pro DAW Ryzen 9 9950X benchmark table (retailer-published) |
| DAWBench VI polyphony, AMD Ryzen 9 9950X, at 64 / 128 / 256 / 512-sample ASIO buffers | 2,140 / 3,780 / 5,660 / 8,580 | Scan Pro Audio, 3XS Pro DAW Ryzen 9 9950X benchmark table (retailer-published) |
| DAWBench DSP plugin instances, Intel Core Ultra 7 270K Plus, at 64 / 128 / 256 / 512-sample ASIO buffers | 512 / 558 / 583 / 618 | Scan Pro Audio, 3XS Project DAW Core Ultra 7 270K Plus benchmark table (retailer-published) |
| DAWBench VI polyphony, Intel Core Ultra 7 270K Plus, at 64 / 128 / 256 / 512-sample ASIO buffers | 2,380 / 4,080 / 6,380 / 9,260 | Scan Pro Audio, 3XS Project DAW Core Ultra 7 270K Plus benchmark table (retailer-published) |
Context note: these are tables published by a system builder on its own product pages, without driver, BIOS or Windows version details, so treat them as indicative rather than independent lab results.
6. Operating-System Audio Stacks: Windows and Android
Before any DAW setting applies, the operating system decides how much buffering sits in the path. Windows 10 cut the shared audio engine’s render latency to 1.3 ms for all applications, from about 12 ms for floating-point apps (Microsoft Learn, Low Latency Audio), yet the default shared-mode buffer stays at about 10 ms unless an app explicitly asks for less. That default is why ASIO, the driver model Steinberg created in 1999, remains the standard for music software on Windows (Microsoft Windows Music Dev blog, 2024); replying to readers on that post, Microsoft engineer Pete Brown wrote that you typically will not get lower than around 10 ms on the native stack, while ASIO drivers often reach single-digit milliseconds.
Microsoft’s answer is a new in-box USB Audio Class 2 driver with an ASIO interface, developed with Yamaha and Qualcomm, with a public preview expected in 2026 through Windows Insider Canary builds (Microsoft, Fall 2025 Windows Musician Technology update). Android publishes harder pass or fail numbers than any desktop OS: the Android 16 Compatibility Definition Document sets ceilings by feature flag, and the Android developer guide admits musicians generally require 10 ms even though its pro-audio flag is defined at 20 ms.
| Metric | Value | Source |
|---|---|---|
| Windows audio engine render latency before Windows 10, floating-point apps vs integer apps | About 12 ms vs about 6 ms | Microsoft Learn, Low Latency Audio |
| Windows 10 and later audio engine render latency | 1.3 ms for all applications | Microsoft Learn, Low Latency Audio |
| Default Windows shared-mode buffer when an app does not request a smaller one | About 10 ms | Microsoft Learn, Low Latency Audio |
| Android continuous output latency for devices declaring the low-latency feature | 45 ms or less | Android Developers, Audio Latency guide |
| Android round-trip latency ceiling, devices declaring FEATURE_AUDIO_PRO | 25 ms (20 ms on an analog path); mean absolute deviation 5 ms | Android 16 Compatibility Definition Document |
| Android round-trip latency ceiling, devices declaring FEATURE_AUDIO_LOW_LATENCY | 50 ms; mean absolute deviation 10 ms | Android 16 Compatibility Definition Document |
| Android tap-to-tone latency ceiling, baseline handheld vs pro audio devices | 250 ms vs 80 ms | Android 16 Compatibility Definition Document |
| Android cold output and cold input latency ceilings, pro audio devices | 200 ms each | Android 16 Compatibility Definition Document |
Context note: mean absolute deviation measures how much the round-trip figure wanders between measurements, which is the CDD’s stand-in for jitter. Tap-to-tone latency is the delay from touching the screen to hearing the resulting sound.
7. Before the Buffer: Controllers, MIDI and Wireless Links
A perfect buffer setting cannot rescue a slow controller path. In the NIME 2016 measurements by McPherson, Jack and Moro, a Bluetooth LE sensor link averaged 139 ms from action to sound, with 1% of tests reaching 334 ms, while the Bela embedded platform held about 0.6 ms with jitter around 20 microseconds. The authors concluded that almost none of the common microcontroller-to-computer setups they tested met accepted latency and jitter guidelines.
USB MIDI fared better than USB serial: a Teensy board sending USB MIDI into Max/MSP averaged 5.1 ms, while an Arduino Uno over USB serial averaged 8.2 ms and exceeded 10 ms 12% of the time. Professional digital wireless sits in a different class, with Shure citing about 2.9 ms for its ULX-D system. The same chain-of-delays problem runs through gaming hardware, as our game latency statistics show for input lag and displays.
| Metric | Value | Source |
|---|---|---|
| Bluetooth LE (BLEmini) controller link, action to sound | 139 ms mean; 1% of tests at 334 ms | McPherson et al., NIME 2016 |
| Arduino Uno over USB serial into Max/MSP, best case | 8.2 ms mean; over 10 ms 12% of the time | McPherson et al., NIME 2016 |
| Teensy over USB MIDI into Max/MSP, best configuration | 5.1 ms mean; 1.5 ms jitter | McPherson et al., NIME 2016 |
| Bela embedded platform, analog sensor input to audio output | About 0.6 ms; jitter around 20 microseconds | McPherson et al., NIME 2016 |
| Digital wireless microphone system latency (ULX-D) | About 2.9 ms | Shure Insights 2016 (vendor’s own claim) |
Context note: McPherson et al. (2016) is the most recent available data for this side-by-side controller comparison. Jitter here means the spread between the shortest and longest delays, which matters as much as the average, as section 1 showed.
Summary: DAW Buffer Size and Processing Latency by the Numbers
| Metric | Value | Source |
|---|---|---|
| Round-trip latency at a 64-sample vs 1,024-sample buffer, 96 kHz ASIO interface | 2.30 ms vs 12.30 ms | MathWorks, Measure Audio Latency example |
| Mean just-noticeable added latency from a 0 ms baseline | 48.8 ms | Schmid et al., Audio Mostly 2024 |
| Mean just-noticeable added latency from a 64 ms baseline | 27.1 ms | Schmid et al., Audio Mostly 2024 |
| Quality rating: steady 10 ms vs zero latency | Not rated lower than zero | Jack et al., Music Perception 2018 |
| Acceptable latency for live performers, by instrument and monitoring | 1.4 ms to 42 ms | Lester and Boley, AES Convention Paper 7198, 2007 |
| Speech rate with immediate vs 200 ms delayed feedback | 5.2 vs 4.3 syllables per second | Ozker and Hagoort, PLoS One 2025 |
| Dysfluencies at 200 ms delay vs shorter delays | Two to three times more | Stuart et al., JASA 2002 |
| Ensemble tempo sweet spot (one-way delay), and share of performances that slowed at 14 ms and above | 11.5 ms; 85% | Chafe and Gurevich, AES 117th Convention 2004 |
| Latency range supporting synchronous ensemble playing | 10 to 25 ms | Washburn et al., Frontiers in Psychology 2021 (review of prior studies) |
| Virtual-instrument polyphony, Ryzen 9 9950X, 64 vs 512-sample buffer | 2,140 vs 8,580 | Scan Pro Audio, DAWBench VI |
| Plugin instances, Ryzen 9 9950X, 64 vs 512-sample buffer | 482 vs 618 | Scan Pro Audio, DAWBench DSP |
| Windows audio engine render latency, Windows 10 and later | 1.3 ms | Microsoft Learn, Low Latency Audio |
| Default Windows shared-mode buffer | About 10 ms | Microsoft Learn, Low Latency Audio |
| Android pro audio round-trip ceiling | 25 ms (20 ms analog) | Android 16 Compatibility Definition Document |
| Bluetooth LE controller link, action to sound | 139 ms mean | McPherson et al., NIME 2016 |
| Bela embedded platform, input to output | About 0.6 ms | McPherson et al., NIME 2016 |
| Recommended in-ear monitor latency | 5 ms and below | Shure Insights 2016 (vendor guidance) |
Methodology and Sources
- MathWorks, Audio Toolbox “Measure Audio Latency” example: loopback round-trip measurements by buffer size at 96 kHz.
- Schmid, Ambros, Bogon and Wimmer, “Measuring the Just Noticeable Difference for Audio Latency”, Audio Mostly 2024: JND by base latency, 37 recruited and 34 analyzed participants.
- Jack, Mehrabi, Stockman and McPherson, “Action-sound Latency and the Perceived Quality of Digital Musical Instruments”, Music Perception 2018.
- McPherson, Jack and Moro, “Action-Sound Latency: Are Our Tools Fast Enough?”, NIME 2016: controller-chain measurements and the Wessel and Wright under-10 ms guideline.
- Lester and Boley, “The Effects of Latency on Live Sound Monitoring”, AES Convention Paper 7198 (2007): range read from the summary in Schmid et al. 2024 because the AES E-Library page could not be accessed.
- Ozker and Hagoort, “Susceptibility to auditory feedback manipulations and individual variability”, PLoS One 2025.
- Ozker, Giglio, Beyh, Forkel and Hagoort, “Individual differences in speech monitoring”, PNAS 2026.
- Stuart, Kalinowski, Rastatter and Lynch, “Effect of delayed auditory feedback on normal speakers at two speech rates”, JASA 2002.
- Chafe and Gurevich, “Network Time Delay and Ensemble Accuracy: Effects of Latency, Asymmetry”, AES 117th Convention 2004.
- Washburn, Wright, Chafe and Fujioka, “Temporal Coordination in Piano Duet Networked Music Performance”, Frontiers in Psychology 2021.
- Scan Pro Audio, 3XS Pro DAW Ryzen 9 9950X benchmark table and 3XS Project DAW Core Ultra 7 270K Plus benchmark table: DAWBench DSP and VI results, accessed October 2026.
- Microsoft Learn, Low Latency Audio (Windows drivers), updated 2025.
- Microsoft Windows Music Dev blog, “Make Great Music with Windows on Arm”, October 2024 and Fall 2025 Windows Musician Technology and Arm64 Update, November 2025.
- Android 16 Compatibility Definition Document, sections 5.6 and 5.10 and Android Developers, Audio Latency guide.
- Ableton Help Center, Driver Error Compensation FAQ and Apple Support, Manage input monitoring latency in Logic Pro for Mac (2026).
- Shure Insights, “Digital Wireless Latency Explained”, November 2016: manufacturer guidance, labeled as such wherever cited.
- Data watch: the Android developer guide defines the pro-audio flag at 20 ms round trip, while the Android 16 CDD table sets 25 ms on most paths and 20 ms only on analog; we report the CDD as the binding document. The CDD’s baseline handheld round-trip row lists two unlabeled values (250 and 200 ms), so we left it out. The Scan Pro Audio numbers are retailer-published product-page benchmarks without driver or firmware details. The MathWorks example does not name its operating system (ASIO points to Windows) and covers one interface, so absolute overhead will differ on other hardware. The Pete Brown “around 10 ms” remark comes from his reply in the comments of a Microsoft blog post, not from the post body. Several perception figures are the most recent available data but are old: Stuart et al. (2002), Chafe and Gurevich (2004), Lester and Boley (2007), McPherson et al. (2016) and the Shure guidance (2016). Microsoft’s in-box ASIO and USB Audio Class 2 driver had no published latency measurements as of its November 2025 update.
Last updated: October 24, 2026. We update this roundup quarterly, and the next refresh is expected when Microsoft releases the public preview of its in-box ASIO and USB Audio Class 2 driver through Windows Insider Canary builds.