Rhythm test — how steady is it once the beat stops?

A browser adds two delays to this measurement that it cannot know to the millisecond: the audio path holds the click back after the page schedules it, and the input path holds your keypress back before the page sees it. A human offset sits on top of both — tappers land ahead of the beat, by 20-60 ms in Repp (2005), Sensorimotor synchronization: a review of the tapping literature, Psychonomic Bulletin & Review, less with musical training. Add the three and an average asynchrony is mostly a number about a laptop. The standard deviation ignores all of them: a constant added to every tap moves the mean and leaves the spread exactly where it was. The page still prints the output latency the browser reports, because a floor that is stated can be argued with.

The clicks are thirty milliseconds of a 1200 Hz square wave, with the count-in accented at 1800 Hz — a short, bright onset is something a finger can aim at, and a soft click widens every figure on the page. All of them are scheduled up front on the audio hardware’s own clock rather than fired from a timer, since a browser timer wobbles by the same order as the quantity being measured. Each tap is matched to its nearest click, so one missed tap does not shift the rest; a tap within a quarter of an interval of the last is read as a bounced key, and a continuation interval outside 40 to 210% of the target as a missed or doubled tap. All three are counted and named.

While the metronome plays, two things run at once: a timekeeper producing intervals and a correction process nudging each tap toward the click. Take the click away and correction has nothing to work with, so the intervals that keep coming are the timekeeper’s own. They drift, and the result reports the drift and the average interval against the target. Timing variability grows with the interval being timed — the smallest change in an interval people can detect is 5-10% of it, per Grondin (2010), Timing and time perception: a review of recent behavioral and neuroscience findings — so the spread is also given as a share of the interval, which makes the four tempos comparable.

The clock-and-finger split follows Wing and Kristofferson’s 1973 model: a tap happens when an internal interval runs out, plus a motor delay, and the two noises are independent. A motor delay that runs long shortens one interval and lengthens the next, so neighboring intervals should be negatively correlated — and the size of that correlation gives the motor part, with the remainder belonging to the clock. When the correlation comes out positive, the model has failed for that run, usually because the run drifted, and the page says so instead of taking the square root of a negative variance.

This is a measurement exercise, not a clinical assessment. It reports what you did on this page against a stated reference and nothing more — it cannot establish a timing or movement disorder. Only a qualified professional, working with more than a browser, can make that judgment.

Every figure here describes one run on one machine, and the machine is inside the number: a slow keyboard scan, a large audio buffer or a click too quiet to give the ear a sharp target each widen a spread that other hardware would call narrow. The comparison worth making is between two of your own runs, taken the same way.