Toward Professional Distributed Performances: Effects of a Global Metronome on Networked Musical Ensemble Interactions

When musicians play together over a network, each one hears the others late. Earlier studies, mostly hand-clapping in pairs, found that delays of 10 to 25 ms give the best synchrony with a stable tempo, that between 25 and 65 ms the tempo tends to slow down, and that beyond that the performance falls apart. One proposed fix is the Global Metronome of Oda and Fiebrink (2016), which is also in this library: every site computes the same click from its own well-synchronized clock, so all clicks line up without being sent over the network.

This JAES 2021 paper by Robert Hupke, Marcel Nophut, Stephan Preihs and Jürgen Peissig (Leibniz University Hannover) is the last step of a series by the same group, run inside the German LIPS project (Live Interactive PMSE Services). Their earlier studies tested the metronome in the lab with a drum pad rhythm and with Djembe duets. Here they move it into a real distributed setup: a rock/pop ensemble of five musicians, three in Hanover (synthesizer, drums, guitar) and two in Munich (grand piano, electric bass), 500 to 600 km apart, linked by a dedicated fiber line, with video walls showing the other room. They compare no metronome, an audible click and a new visual LED metronome, and measure both what the recordings show and what the musicians felt.

• The global metronome inside a professional NMP. The authors describe this as the first time an auditory and a visual global metronome were integrated into a physical networked performance, rather than a lab setup.
• PTP instead of GPS. The metronome (one Raspberry Pi per site, called GloMe) was extended to follow the Precision Time Protocol that already synchronized every camera, audio interface and AD/DA converter in the project. Because the Raspberry Pi has no hardware time stamping, they implemented software time stamping. Two PTP-synchronized metronomes produced audio clicks with a standard deviation of 92 μs between them over 30 minutes. For comparison, their earlier GPS version reached 56 μs. Once both devices share a clock, each one computes and plays the next beat locally, as in Oda and Fiebrink; no click is sent between sites.
• A visual metronome. After talking to the musicians they added an LED board of eight light modules in a small housing, powered from the Pi's 5 V supply and driven from a single GPIO pin by a bitstream carrying a color value per module. Light flash and audio click differed by a standard deviation of 66 μs over 30 minutes, so the light is effectively on the same clock as the sound.
• An ecologically valid test with three conditions (no metronome, acoustic, visual), three latencies and three tempos, analyzed with both objective metrics and a questionnaire.

For a software engineer, the control plane is the interesting part. The researcher's workstation in Hanover set the tempo of both metronomes and started them over SSH. The extra network delay and the fade out of the click were controlled over UDP messages to the mixing consoles in both cities. So the only things a remote command changes are the parameters (tempo, start, on or off); the beat itself comes from the shared PTP time base at each site. The paper does not spell out how the two Pis agree on the beat phase (for example, a common start instant or the epoch based formula of Oda and Fiebrink); it points to the group's earlier implementation papers for that.

• Network and audio. A rented point-to-point fiber link carried six audio channels; the drums were pre-mixed in Hanover and sent to Munich as a stereo mix. Everything was recorded in Hanover, including the click from both sites as a reference. Musicians wore in-ear monitors (Sennheiser IE400 PRO) with their own mixes; in Munich they switched to open headphones (Sennheiser HD800) after the second series. A talk-back microphone in each room was used between trials.
• Latency levels. The measured end-to-end baseline latency was 8.5 ms. Extra delay was added at the outputs of the mixing consoles feeding the remote site, giving three levels: 8.5, 25.5 and 55.5 ms (reported as one-way delay). Self-delay, used in their earlier work, could not be implemented because of mixing console limits.
• Tempo levels. 90, 105 and 144 bpm.
• Conditions. Every trial started with an eight-beat audible count-in on the metronome, with all musicians starting together (no Leader/Follower roles). In condition 1 the click was muted by the end of the second measure but still recorded. In condition 2 the click stayed on for the whole trial. In condition 3 the audio click was faded out the same way, while the LED metronome was on from the count-in to the end.
• Design. 27 trials (every combination of delay, tempo and condition) in random order, repeated on four days within one month, for about 120 minutes per session. With five instruments and two metronome tracks this gave 756 recordings.
• The piece. Written by the musicians themselves: a four-measure sequence in 4/4, repeated at least four times, about 50 s long. Each part has a different rhythm, but all meet at the start of each measure.
• The musicians. Students of the music universities in Hanover and Munich. The two groups had never met before; within each city they knew each other but had little experience playing together. The authors say the technical effort and short time window did not allow more musicians.
• Signal analysis. Onsets were found with MIRtoolbox: a 10-band Gammatone filter bank, differentiated and half-wave rectified envelopes, and selected bands summed into a novelty curve per instrument. The metronome and bass drum peaks served as reference points for the first note of each measure. A single missing note was linearly interpolated, which happened in 20 trials; no trial was discarded.
• Metrics. From the inter-onset intervals between measure starts they computed tempo in bpm, smoothed it with a three-point moving average and fit a linear regression, reported as relative tempo deviation in percent. Synchrony was measured as asymmetry, the mean time difference between two players at the synchronization points, corrected by the round-trip delay for the remote perspective.
• Questionnaire. After each trial, Likert ratings on interplay quality, effort to stay in sync with remote and local players, stress, and (with a metronome) self-confidence and use of the metronome, plus free comments. A second questionnaire at the end asked for general feedback.
• Statistics. Shapiro-Wilk tests showed that not all conditions were normally distributed, so they used Friedman's test per tempo and delay, with Tukey HSD post hoc comparisons.

Values are the authors' reported means of relative tempo deviation (negative means the band slowed down).

• Without a metronome the tempo drifted, in a direction that depended on the starting tempo. At 90 bpm the band sped up at every delay, which goes against the usual expectation of slowing down. At 105 bpm it slowed at the higher latencies, and at 144 bpm it slowed at every delay.
• The acoustic metronome kept the tempo close to 0% deviation with small variance at every delay and tempo, with a slight acceleration at 55.5 ms for all three tempos. The visual metronome showed a similar trend but with more variance.
• Where it helped most. At 8.5 ms and 144 bpm the band went from minus 1.57% without metronome to 0.17% with the acoustic and 0.29% with the visual one. At 25.5 ms and 144 bpm, from minus 0.72% to 0.27% and 0.26%. At 25.5 ms and 90 bpm the baseline was 1.15%, the acoustic 0.16% and the visual minus 0.088%. At 55.5 ms the metronome only had a significant effect at 105 bpm (acoustic better than both no metronome and visual). At 8.5 ms and 90 bpm the change from 0.68% to minus 0.72% with the visual metronome was significant but not an improvement.
• Synchrony depended on the people, not the metronome. For the musicians in Hanover the asymmetry relative to the drummer stayed between minus 32.3 and minus 7.3 ms in all conditions. For Munich, without a metronome the asymmetry roughly equaled the round-trip delay, which means they aligned to the drums they heard. The metronome mainly removed the trend of asymmetry growing with tempo, and the acoustic one brought small improvements at 25.5 and 55.5 ms.
• Strategies varied trial by trial. Looking at the pianist in Munich and the drummer in Hanover, without a metronome the pianist sometimes anticipated by about the round-trip delay so the drummer could play on time, and at higher tempos fell back to following the drummer. With the acoustic click there was no consistent strategy; in some trials both played to their own click, and at 55.5 ms the pianist placed himself halfway between his click and the drums. With the visual metronome the pianist almost always followed the drummer.
• Subjective ratings. Interplay quality dropped and effort rose as latency rose, as expected. The acoustic metronome made no significant difference to either. The visual one was rated significantly worse in interplay quality and significantly higher in effort at 25.5 ms for 105 and 144 bpm. Both metronomes were rated as giving self-confidence at almost every delay and tempo, and as mostly used, more so at 144 bpm.
• Free comments. The drummer ignored the Munich players at high latency to keep the tempo stable, found the click a big help and the LED not always. The Hanover players used the click as a reference when the drummer drifted; the guitarist played behind the beat and the synth player played legato to hide imprecise sync points. Everyone preferred the audible click, reported less frustration over time, and the Munich players suggested a drummer count-in so they could judge the delay against the click.

The engineering part is solid: the clock error (tens of microseconds) is roughly two orders of magnitude below the latencies tested, so any effect comes from the musicians, not from metronome jitter. The musical study is much thinner than its title.

• One ensemble, five people. All 20 data points per tempo and condition come from the same five musicians over four days, and the authors admit they could not include more. These are students, not an established professional band, although the abstract calls them a professional ensemble. With one band, the results mix the metronome effect with this drummer, this pianist and their personal strategies.
• One simple piece. A four-measure loop of about 50 s, written by the band, with sync points only at measure starts. That is good for onset detection and reproducibility but leaves out longer songs, tempo changes and rubato.
• Statistics are fragile. Friedman's test is run separately for each of nine delay and tempo cells and for several metrics with no correction for multiple comparisons, and the repeated measures from the same five people are not independent. P values seem to be floored at 0.01: χ²(2) = 21.9 corresponds to a p value far below 0.001, yet it is printed as 0.01. At 55.5 ms and 90 bpm the test gives χ² = 0 and all pairwise differences are exactly zero, which is odd and not explained.
• Text and tables disagree. For tempo at 25.5 ms and 90 bpm, the text says the significant differences are baseline vs acoustic and visual vs acoustic, but Table 4 shows baseline vs acoustic at p = 0.61 and baseline vs visual at p = 0.01. For Hanover asymmetry at 8.5 ms and 90 bpm the text gives χ² = 6.5, p = 0.02, while Table 5 gives 6.25 and 0.04. Section 2.3 calls the post-trial questionnaire eight questions and Section 4.2 says seven; Table 1 lists eight items, and Q2.3 (which signal they focused on) is never reported. Table 10 also shows a significant effort effect at 8.5 ms and 105 bpm that the text skips.
• Abstract and conclusion overstate. The abstract says the metronome helps tempo stability at high latency, but most significant tempo effects are at 8.5 and 25.5 ms, and at 55.5 ms only 105 bpm is significant. The conclusion says neither metronome affected the effort to stay in sync, but the results show the visual one was rated significantly more effortful at 25.5 ms. The conclusion also says the high subjective usefulness of the visual metronome is not matched by the objective data, while the discussion says the visual metronome achieved nearly equally good objective results. The body figures are the ones to trust.
• A confound in the setup. The Munich players changed from in-ear to open headphones after the second series, which changes what they hear, and it is not analyzed.
• The network is a best case. A dedicated point-to-point fiber with PTP across the whole infrastructure is far from a home or office Internet link. Latency here is a fixed added delay, without jitter or packet loss. The paper also does not say where the PTP grandmaster sits or how PTP behaves over 500 to 600 km, which is exactly what an implementer would want to know. Over the public Internet, GPS (or a local GPS-disciplined server per site) is the safer clock source, and the group's earlier GPS figure of 56 μs supports that.
• The visual metronome is a blink, not a conductor. As the authors note, the LED shows no "beat coming", unlike a baton. A design with a countdown or a sweeping indicator might change the visual results completely.

For someone building this, the takeaway is still clear: synchronize clocks (PTP or GPS), compute the click locally at each site, send only tempo and start commands, and expect the shared click to stabilize tempo but not to fix how people react to late audio.

• Larger samples and more ensembles, to separate the metronome effect from individual musicians' strategies, which the authors name as necessary.
• Deeper study of musicians' attention and focus during ensemble interaction with a metronome.
• Easier control of the metronome by the musicians themselves, for example haptic control, instead of a researcher over SSH.
• An adaptive metronome that follows the players' tempo, and a visual mapping of a global conductor, both cited by the authors as next steps.
• A visual metronome that shows the beat approaching, a self-delay condition, and tests over shared Internet links with jitter, which this setup could not cover.