Ensemble Timing and Latency: Why 25 Milliseconds Matters
Two people playing together are a coupled system with a delay in the loop. Move them apart in time and the tempo itself changes, in a direction that depends on how far apart they are.
Short answer: every ensemble already plays with latency, because sound crosses a room at about 343 metres per second. A drummer three metres from a bass player hears them roughly 9 milliseconds late. That is normal, and research suggests it is close to the range in which people synchronise best. Push the delay much below it and duos tend to speed up. Push it well above it and they slow down, then fall apart. The figure usually quoted as the working limit for playing together over a network is about 25 milliseconds of one-way delay.
What an ensemble is doing when it plays in time
Keeping time together is not a matter of each player independently following an internal clock. It is mutual correction. Every player hears the others slightly late, compares that against their own last note, and adjusts the next one. The adjustment is small and mostly unconscious, and it happens on every beat.
Wing, Endo, Bradbury and Vorberg made this concrete for a string quartet in the Journal of the Royal Society Interface in 2014. Modelling the note onsets of two professional quartets, they estimated how strongly each player corrected towards each of the others. The correction was not symmetric. In one quartet the first violin corrected less towards the others than they corrected towards the first violin, which is the mathematical signature of a leader. In the other quartet the correction was spread more evenly. Both quartets played in time. They simply distributed the timing work differently.
That is the useful frame for everything below. Latency does not just make you late. It corrupts the feedback the correction depends on, and a correction loop with too much delay in it stops behaving like a correction loop.
What delay actually does to a duo
Chafe, Cáceres and Gurevich tested this directly and published the result in Perception in 2010. Pairs of participants clapped a rhythm together from separate sound-isolated rooms, through headphones, with no visual contact. The experimenters inserted one-way delays between the pair ranging from 3 to 78 milliseconds and measured what happened to the performance.
The finding is more interesting than a simple tolerance limit. There were distinct regimes of coupled behaviour:
- Very short delays, shorter than any real physical separation would produce, tended to make the pair accelerate. The authors attributed this to anticipation: with almost nothing in the loop to absorb it, each player's tendency to arrive slightly early compounds.
- Natural delay, the narrow band matching the travel times you get from ordinary spatial arrangements of an ensemble, supported the most stable performance.
- Longer delays produced an increasingly severe deceleration, and then outright deterioration of the performed rhythm.
So the room you rehearse in is not an obstacle to be engineered away. The few milliseconds of air between players appear to be part of what makes the coupling stable. That connects directly to what our guide to anticipation versus reaction covers: players habitually aim slightly ahead of where they think the beat is, and the acoustic delay of the room absorbs part of that tendency.
The 25 millisecond figure, and where it comes from
Anyone who has read about playing music over the internet has met the number 25. In the networked music performance literature it has a name: the ensemble performance threshold, the maximum end-to-end delay under which musicians can still play in synchronisation as if they were in the same place. Rottondi, Chafe, Allocchio and Sarti set it out in their review of networked music performance technologies in IEEE Access in 2016, alongside a survey of the perceptual studies the figure rests on.
It is worth converting that into distance, because it makes the number intuitive rather than arbitrary. Sound covers roughly 34 centimetres per millisecond in air. Twenty-five milliseconds of delay is therefore about 8.5 metres of separation, which is a wide stage but not an absurd one. Orchestras deal with more than that, and they deal with it using a conductor, sightlines and rehearsed convention rather than by relying on the sound arriving on time.
Over a network the same 25 milliseconds buys far less. Light in optical fibre travels at roughly two thirds of its speed in vacuum, so the theoretical floor is about one millisecond per 200 kilometres in each direction, before any equipment is involved. In practice, routing, buffering and analogue conversion consume most of the budget long before the distance does.
The delay budget, item by item
| Source of delay | Roughly how much | Can you reduce it? |
|---|---|---|
| Air between two players, 3 m apart | About 9 ms | Stand closer, or accept it |
| Air across an 8.5 m stage | About 25 ms | Monitoring, sightlines, a conductor |
| Audio interface buffer, 128 samples at 48 kHz | 2.7 ms per buffer, and there is more than one | Smaller buffer, at the cost of stability |
| Fibre path, 500 km each way | About 2.5 ms one way at best | No, this part is physics |
| Consumer video-call software | Typically far above the ensemble threshold | Use software built for the job |
| Bluetooth headphones | Tens of milliseconds, sometimes far more | Use wired monitoring for any timing work |
The last row catches most people. A wireless audio link adds enough delay on its own to put you outside the ensemble threshold before the network is involved at all. If you are measuring your own timing rather than playing with someone else, our audio latency and timing offset test exists to find out how much your own chain is adding.
How much can musicians work around?
The 25 millisecond figure is the threshold for playing as though you were together. It is not the point where music becomes impossible. Bartlette, Headlam, Bocko and Velikic explored the wider range in Music Perception in 2006. Two pairs of musicians played Mozart duets while visually isolated and connected through microphones and headphones, under delays of 0, 20, 40, 50, 80, 100, 120, 150 and 200 milliseconds. Each musician heard themselves in real time, with only the other part delayed.
Two things stand out. First, the musicians adopted different strategies for coping, and those strategies produced measurably different degrees of success at holding coordination, pacing and regularity. Second, whatever the strategy, both duets were strongly affected at and above 100 milliseconds.
Between roughly 25 and 100 milliseconds, then, you are in a region where playing together is possible but the ensemble is doing extra work, and where who leads becomes a decision rather than an accident. That fits what Goebl and Palmer reported in Music Perception in 2009 for pianist duos: as the auditory information each player had about the other was reduced, synchronisation degraded, and the players leaned harder on other channels, including visible body movement.
What to do about it
Stop treating the rehearsal room as neutral. Where people stand changes the delays inside the correction loop. A drummer who cannot hear the bass directly, only through a monitor wedge, is working with a delay their ears did not predict.
If you play over a network, measure before you blame the players. A duo that keeps slowing down is behaving exactly the way the 2010 clapping experiment predicts for excessive delay. That is a diagnosis, not a criticism of anyone's time.
Remember that recording to a click has latency too. The click you hear is delayed by your monitoring chain, and your playing is captured after a further delay. The two do not have to cancel out, which is why offsets are worth measuring rather than assuming.
A note on the games here. Chronkle's tempo tasks are single player, so they measure your timing against a reference rather than against another person, which deliberately removes the coupling this article is about. The one thing they do share with ensemble playing is how quickly timing drifts once the reference goes away.
Hold a tempo while something pushes against it. Chrondrift asks you to keep a drifting tempo inside a target band. It is the solo version of the problem an ensemble solves collectively, and it shows very quickly how a small correction error compounds.
Frequently asked questions
What is the maximum latency for playing music together online?
The figure used in the networked music performance literature is about 25 milliseconds of end-to-end delay, known as the ensemble performance threshold. Rottondi, Chafe, Allocchio and Sarti describe it in their 2016 IEEE Access review. Above that, playing as though you were in the same room stops working reliably, although music is still possible with adjusted strategies up to around 100 milliseconds.
Does a small delay always make an ensemble slow down?
No, and this is the surprising part. Chafe, Cáceres and Gurevich found in Perception in 2010 that very short delays, shorter than any real physical separation, tended to make clapping duos accelerate, apparently because anticipation is fed back with almost nothing to absorb it. Deceleration appeared at delays longer than natural room distances, and got worse as the delay grew.
How far apart is 25 milliseconds in a room?
Sound travels roughly 34 centimetres per millisecond in air, so 25 milliseconds is about 8.5 metres. That is a wide stage. Ensembles spread over that distance rely on sightlines, monitoring and a conductor rather than on the sound arriving in time.
At what delay do musicians really struggle?
Bartlette, Headlam, Bocko and Velikic tested delays up to 200 milliseconds with musicians playing Mozart duets, published in Music Perception in 2006. Coordination, pace and timing regularity were strongly affected at and above 100 milliseconds, though different players found different coping strategies with different degrees of success.
Do Bluetooth headphones matter for this?
Yes, more than most people expect. A wireless audio link commonly adds tens of milliseconds on its own, which can exceed the whole ensemble budget before any network is involved. For any playing or measuring where timing matters, use a wired connection.
Why does a real room feel better than a perfectly synchronised link?
Because the small acoustic delay of a room appears to be part of what stabilises mutual correction. In the 2010 clapping study, the narrow band of delay corresponding to normal ensemble spacing supported the most stable performance, and conditions on both sides of it produced characteristic instabilities.
Keep reading
- Anticipation versus reaction: why players aim slightly ahead of the beat, and what absorbs that.
- Rushing and dragging: the same acceleration and deceleration, from causes inside the player.
- Sensorimotor synchronisation: the tapping research the ensemble models are built on.
Sources: Chafe, C., Cáceres, J.-P. and Gurevich, M. (2010), Effect of temporal separation on synchronization in rhythmic performance, Perception; Rottondi, C., Chafe, C., Allocchio, C. and Sarti, A. (2016), An overview on networked music performance technologies, IEEE Access; Bartlette, C., Headlam, D., Bocko, M. and Velikic, G. (2006), Effect of network latency on interactive musical performance, Music Perception; Wing, A. M., Endo, S., Bradbury, A. and Vorberg, D. (2014), Optimal feedback correction in string quartet synchronization, Journal of the Royal Society Interface; Goebl, W. and Palmer, C. (2009), Synchronization of timing and motion among performing musicians, Music Perception.