# Finding the beat

A field guide to rhythm tracking for audiovisualizers. Updated 21 November 2026.

[Interactive edition](https://research.klob.io/)

## Rhythm in layers

Read the same waveform as a sequence of sounds, a pulse, a bar, and a musical phrase.

A repeated kick can tell you when to brighten a shape. A visual rotation that completes on every bar needs a clock that keeps moving through a missing kick. A scene change on the next drop needs information about the arrangement. Those three effects can look at exactly the same recording and need different answers. [[1]](https://cs-web.bu.edu/fac/snyder/cs583/Literature%20and%20Resources/bello_onset_tutorial.pdf) [[11]](https://www.eurasip.org/Proceedings/Eusipco/Eusipco2006/papers/1568981125.pdf) [[17]](https://mir-aidj.github.io/raveform/)

### Five things a listener can hear

| Target | What to return | Example at 124 BPM in 4/4 |
|---|---|---|
| Onset | A timestamp and strength for an acoustic attack | An offbeat hi-hat arrives about 242 ms after the kick. |
| Tempo | The rate of a chosen pulse, globally or over time | 124 quarter-note beats per minute. |
| Beat | A timestamped sequence, or a continuously advancing phase | One quarter note every 483.87 ms, including silent pulses. |
| Downbeat | The beat that opens each measure | Every fourth beat, once bar phase is established, about 1.94 s apart. |
| Phrase or section boundary | A timestamp, musical level, and ideally a cue type | If the arrangement groups eight bars together, those group starts are 15.48 s apart. |

A melodic phrase can start with a pickup before beat one. A functional section such as a chorus or drop may contain several phrases. A DJ cue marks a useful performance choice and can be placed before either. Decide which kind of timestamp a visual expects before choosing a detector or reference dataset. [[18]](https://arxiv.org/abs/2611.13503) [[17]](https://mir-aidj.github.io/raveform/) [[20]](https://arxiv.org/abs/2407.06823)

### Read the onset envelope

The cheapest detector measures short-time energy or root-mean-square amplitude. Compare each frame with a running baseline, trigger on a rise, and wait briefly before accepting another event. A low-frequency version can follow a kick-heavy mix. A sustained synth raises the baseline, a snare roll produces many triggers, and a silent beat leaves no peak. As a visual input it still works well for a momentary pulse or a change in glow. [[1]](https://cs-web.bu.edu/fac/snyder/cs583/Literature%20and%20Resources/bello_onset_tutorial.pdf) [[30]](https://github.com/projectM-visualizer/projectm/wiki/Preset-Authoring-Guide)

A more selective detector windows the samples and takes a short-time Fourier transform (STFT). Compare adjacent magnitude spectra and sum their positive changes. This **spectral flux** can expose a new hi-hat while the bass stays loud. Use log compression so loud bins dominate less, normalize recent history, and pick local peaks above an adaptive threshold. Separating bass, middle, and treble lets a kick and a vocal consonant carry different weights. [[1]](https://cs-web.bu.edu/fac/snyder/cs583/Literature%20and%20Resources/bello_onset_tutorial.pdf) [[2]](https://www.dafx.de/paper-archive/details/0oee-99Z88WL7pSo749gcA)

Complex-domain detectors also compare observed Fourier coefficients with predictions from recent magnitude and phase. SuperFlux takes a different route, allowing energy to move between nearby frequency bins when comparing frames. This suppresses false onsets from vibrato. A bowed note with a slow attack can still have an uncertain onset time even for a human annotator. [[1]](https://cs-web.bu.edu/fac/snyder/cs583/Literature%20and%20Resources/bello_onset_tutorial.pdf) [[2]](https://www.dafx.de/paper-archive/details/0oee-99Z88WL7pSo749gcA)

### Choose an analysis clock

An illustrative browser configuration uses 48 kHz audio, a 1,024-sample window and a 256-sample hop. Each window covers 21.33 ms and a new feature arrives every 5.33 ms. A centered window needs about 10.67 ms of future samples for the timestamp at its center. A trailing window is causal, but its evidence still takes time to accumulate. Feature timestamps, decision times, and the time a pixel lights up are different quantities. The values here are design examples. [[36]](https://developer.mozilla.org/en-US/docs/Web/API/AudioContext/getOutputTimestamp) [[1]](https://cs-web.bu.edu/fac/snyder/cs583/Literature%20and%20Resources/bello_onset_tutorial.pdf)

## Estimating tempo

Compare thresholds, interval histograms, resonators, and local tempograms.

Tempo describes how quickly a selected pulse repeats. Its period in seconds is 60 divided by BPM. A 124 BPM pulse and a 62 BPM pulse can both fit the same house track if one follows the kick and the other follows every second kick. A tempo estimator must choose the pulse level, and a visualizer must decide what to do when that choice is uncertain. [[37]](https://transactions.ismir.net/articles/10.5334/tismir.43) [[4]](https://www.audiolabs-erlangen.de/resources/MIR/FMP/C6/C6S2_TempogramAutocorrelation.html)

### Tap tempo and interval histograms

A manual baseline takes the median interval between recent taps and converts it to BPM. Display how widely the intervals vary and let a performer tap again after a transition. This also provides a reference when judging machine estimates by ear.

For a machine-only baseline, detect transients and collect their inter-onset intervals. A histogram of adjacent gaps works on a plain click track. Including a limited set of nonadjacent pairs can recover the beat despite subdivisions or missing notes. Cluster similar ratios or vote for several multiples of a gap. Keep the original tempo hypotheses so a tempo prior can be changed without losing the evidence. A loud eighth-note hat naturally creates a double-time candidate. [[5]](https://webspace.eecs.qmul.ac.uk/s.e.dixon/pub/2007/jnmr07.pdf) [[37]](https://transactions.ismir.net/articles/10.5334/tismir.43)

### Autocorrelation, resonators, and tempograms

For an onset-strength sequence and each lag, sum the products of values separated by that lag. Peaks in this **autocorrelation** indicate recurring periods. At a feature rate of 100 Hz, a lag of 48 frames corresponds to 125 BPM. Interpolation or fitting over longer spans improves resolution. The strongest peak can be a multiple of the beat period. A genre-aware tempo range or bias can help choose between candidates, but a hard range can force a convincing wrong answer. [[6]](https://www.ee.columbia.edu/~dpwe/pubs/Ellis07-beattrack.pdf) [[4]](https://www.audiolabs-erlangen.de/resources/MIR/FMP/C6/C6S2_TempogramAutocorrelation.html)

Scheirer used multiple frequency bands and banks of resonating comb filters. A filter tuned to a repeating interval builds up a response when accents recur at that interval. Separate frequency bands allow rhythmic agreement to survive a change in instrumentation. A resonator needs time to build evidence and can keep ringing when the tempo changes. [[3]](https://pubmed.ncbi.nlm.nih.gov/9440344/)

A **tempogram** repeats a periodicity analysis in sliding windows. An autocorrelation tempogram expresses periodicity by lag. A Fourier tempogram measures modulation frequencies in the onset envelope and can show harmonics of the fundamental pulse rate. Long windows separate nearby tempi but blur transitions. Short windows react faster and have fewer repetitions to distinguish a beat from its subdivisions. Plotting several candidates makes a tempo ramp or half-time switch easier to diagnose than displaying one rounded BPM value. [[4]](https://www.audiolabs-erlangen.de/resources/MIR/FMP/C6/C6S2_TempogramAutocorrelation.html)

### Constant tracks, live drummers, and blends

A sequenced track is often well served by a constant period and a fitted offset. For a live drummer, estimate a local tempo trajectory, penalize abrupt jumps, and allow sustained evidence to move the estimate. In a DJ crossfade there may be two plausible pulse trains. For the visual clock, favor the currently locked train until a second candidate wins consistently or the operator switches decks. This proposed control policy keeps the visual clock stable while both decks are audible. [[5]](https://webspace.eecs.qmul.ac.uk/s.e.dixon/pub/2007/jnmr07.pdf) [[9]](https://www.audiolabs-erlangen.com/resources/MIR/2024-TISMIR-RealTimePLP)

Tempo-CNN makes a useful offline comparison. It classifies tempo directly from 11.9 seconds of mel-spectrogram input. Overlapping excerpts can show tempo drift. The output still leaves the position of the next beat unresolved. At 124 BPM a half-beat offset is about 242 ms, large enough to move a flash from the kick to the offbeat hat. [[22]](https://ismir2018.ircam.fr/doc/pdfs/141_Paper.pdf)

## Tracking the beat

Follow a pulse through missing kicks, syncopation, and changing tempo.

A beat tracker has to answer “when?” as well as “how often?” It can place a beat where the audio has no attack, provided neighboring events establish a pulse. That ability matters when a kick drops out under a vocal or a drummer anticipates the next bar. [[6]](https://www.ee.columbia.edu/~dpwe/pubs/Ellis07-beattrack.pdf) [[9]](https://www.audiolabs-erlangen.com/resources/MIR/2024-TISMIR-RealTimePLP)

### A grid with a phase tracker

For a steady track, test possible offsets across one candidate beat period. Sum onset strength near each predicted grid point and retain the best-supported alignment. During playback, a phase-locked-loop-style tracker can predict the next beat and use nearby evidence to adjust its phase and period. Small bounded corrections keep motion smooth. A large correction should require several consistent observations or a manual re-anchor. Otherwise an isolated fill can move the entire show.

BeatRoot keeps competing hypotheses. After clustering inter-onset intervals, agents with different periods and starting times attempt to match new events. An agent can survive a missing onset and a rival can win after the first lock proves false. This trades extra state and tuning for a way out of an early phase error. [[5]](https://webspace.eecs.qmul.ac.uk/s.e.dixon/pub/2007/jnmr07.pdf)

### Dynamic programming across a file

Ellis scores a beat path using onset strength at its chosen timestamps and a penalty for intervals that depart from a target period. One implementation uses a squared log-ratio penalty, so doubling and halving an interval incur symmetric penalties. Dynamic programming stores the best predecessor at each candidate time and backtracks through the recording. This lets evidence later in the file change an earlier decision. librosa exposes the approach and accepts a supplied BPM estimate, including time-varying values in its documented interface. [[6]](https://www.ee.columbia.edu/~dpwe/pubs/Ellis07-beattrack.pdf) [[7]](https://librosa.org/doc/main/api/generated/librosa.beat.beat_track.html)

A short trailing block can also be decoded, but committing its last beat is a different algorithmic contract. The answer may move when a later block arrives. A live renderer should distinguish provisional history, committed events, and future predictions. A correct timestamp computed after the frame deadline cannot trigger that frame.

### Causal pulse and prediction

BTrack is a classical C++ implementation for beat-synchronous processing. It combines local onset evidence with a tempo model and beat prediction. The 2024 real-time predominant local pulse (PLP) system instead uses local periodic components to construct a continuous pulse. The implementation exposes stability and future pulse peaks, useful for scheduling a motion apex. aubio provides another low-overhead streaming baseline with beat, period, and confidence outputs. [[8]](https://github.com/adamstark/BTrack) [[9]](https://www.audiolabs-erlangen.com/resources/MIR/2024-TISMIR-RealTimePLP) [[10]](https://aubio.org/doc/latest/tempo_8h.html)

An extrapolated beat can drive motion on time, then be contradicted by a sudden stop. Track the last credible observation and decay confidence during gaps. A brief dropout can coast for a musical interval, while a long breakdown can retain smooth motion and suppress sharp kick accents. A seek, loop jump, or newly selected deck should invalidate queued events from the former timeline. These are visual control decisions to test against the performance material.

### A measured phase failure

Vande Veire and De Bie evaluated a specialized steady-tempo drum-and-bass tracker on a second set of 220 tracks in the 160–190 BPM range. It aligned correctly on 216. In all four failures it found the right tempo and placed the entire beat grid half a beat away. A BPM-only score would have concealed all four failures. [[12]](https://biblio.ugent.be/publication/8577611)

## Finding beat one

Resolve which beat opens the measure and recognize when the audio leaves it ambiguous.

The first beat of a measure is the downbeat, often called “the one.” A 4/4 beat tracker can follow all the right kicks yet call beat three the start of every bar. To infer bar phase, the audio must give the tracker evidence that distinguishes at least some of those four positions. [[11]](https://www.eurasip.org/Proceedings/Eusipco/Eusipco2006/papers/1568981125.pdf) [[21]](https://www.cp.jku.at/research/papers/Boeck_etal_ISMIR_2016.pdf)

### Four possible answers in 4/4

Start with a sequence of beats. Group them into bars under each of four offsets. For each offset, compare beat-aligned features across several bars. Useful votes can come from a bass-note or chord change, a kick following a fill, a crash entering, a repeated snare pattern, or a change of instrumentation. Score their consistency over time. Davies and Plumbley modeled this using beat-synchronous spectral differences and bass-frequency evidence with a known meter. [[11]](https://www.eurasip.org/Proceedings/Eusipco/Eusipco2006/papers/1568981125.pdf)

The strongest hit in a bar often belongs to a snare on two or four. A backbeat alone can leave a two-beat ambiguity. An identical four-on-the-floor kick on all four beats has no information identifying beat one. Waiting for a bass or harmonic change, importing a deck marker, or taking a manual “one” can resolve the offset. A count from the first detected beat is only as good as that initial anchor. [[11]](https://www.eurasip.org/Proceedings/Eusipco/Eusipco2006/papers/1568981125.pdf) [[33]](https://support.native-instruments.com/support/solutions/articles/69000879473-how-to-sync-traktor-to-other-applications-via-ableton-link)

### Joint meter and phase inference

A probabilistic tracker can represent tempo, within-bar position, and meter together. The observation model scores evidence for an ordinary beat and a downbeat. The transition model enforces plausible movement through a bar and plausible tempo change. A dynamic Bayesian network (DBN) can select a full-file path with Viterbi decoding. A forward filter or particle filter maintains several possible current states without needing the rest of the song. The decoder, features, and peak picker must all respect the live latency budget. [[21]](https://www.cp.jku.at/research/papers/Boeck_etal_ISMIR_2016.pdf) [[23]](https://arxiv.org/abs/2108.03576)

A 4/4 prior is a sensible default for a dance set. Keep a meter setting or additional hypotheses for 3/4, edits that insert a beat, and material whose compound grouping changes which pulse the listener taps. A missing beat in a rigid counter can rotate the labels on every subsequent bar. Report meter and confidence alongside the most likely bar position. Let the operator reset beat one without changing a tempo that is already correct. [[11]](https://www.eurasip.org/Proceedings/Eusipco/Eusipco2006/papers/1568981125.pdf) [[23]](https://arxiv.org/abs/2108.03576) [[33]](https://support.native-instruments.com/support/solutions/articles/69000879473-how-to-sync-traktor-to-other-applications-via-ableton-link)

### What a transport knows

Ableton Link can share tempo and phase modulo a chosen quantum. Four beats is a useful quantum for four-beat bars. A transport and a correctly analyzed track can supply an excellent clock. The link itself cannot hear which beat the performer intended as the musical downbeat. Native Instruments documents a Reset Downbeat mapping for Traktor when alignment slips after playback resumes. Keep the manual bar anchor independent from the phrase anchor. [[32]](https://ableton.github.io/link/) [[33]](https://support.native-instruments.com/support/solutions/articles/69000879473-how-to-sync-traktor-to-other-applications-via-ableton-link)

## Finding the phrase

Separate a predicted group of bars from a melodic entrance, a drop, or a DJ cue.

When a VJ asks for the next phrase, they might want the next eight-bar mixing point, the start of a vocal line, or the instant the arrangement changes into a drop. Those targets can coincide. They also diverge on pickups, extended builds, and drops that last several phrases. A useful event carries a timestamp and identifies the musical level that produced it. [[17]](https://mir-aidj.github.io/raveform/) [[18]](https://arxiv.org/abs/2611.13503) [[20]](https://arxiv.org/abs/2407.06823)

### Count bars, with an anchor

A simple predictor accepts an operator cue at a known group start, counts downbeats, and projects every fourth, eighth, or sixteenth bar. The length is configurable. This gives advance notice and stays stable through a quiet passage. It will keep counting across an extra two-bar fill, and a mistaken starting bar displaces every later cue. In an unchanging eight-bar loop, several possible starting bars can explain every sample equally well.

Raveform’s EDM annotators used four- and eight-measure organization as a soft prior and followed audible changes elsewhere. Some long drops remained a single labeled section. For an engineering baseline, keep several candidate lengths and offsets, favor recurring agreement, and allow a strong musical event or operator cue to re-anchor the count. Present the countdown as a forecast. [[17]](https://mir-aidj.github.io/raveform/)

### Listen for a change at more than one scale

Average features over beats or bars. Band energies and mel-frequency cepstral coefficients (MFCCs) capture changes in loudness and timbre. Chroma describes pitch-class content and can expose harmonic change. Drum or vocal activity helps identify an entrance. A simple online change detector compares a new bar with a trailing baseline, smooths the score, and requires enough separation between accepted boundaries. A sudden filter sweep can then score highly even while the same phrase continues. [[13]](https://www.audiolabs-erlangen.de/resources/MIR/FMP/C4/C4S4_NoveltySegmentation.html) [[31]](https://personales.upv.es/jjlopez/publications/SegmentEDM_AES144th.pdf)

In Foote-style novelty analysis, build a self-similarity matrix whose cells compare musical frames. At each candidate boundary, a checkerboard kernel rewards similarity within the preceding and following neighborhoods and contrast across them. Vary the kernel width to change the musical scale. A symmetric comparison must hear what comes next. In a drum-and-bass system, combining MFCC and RMS novelty used four measures on either side. At 124 BPM in 4/4, that right-hand window alone takes **7.74 seconds**. This is practical for prepared files and useful for retrospective confirmation. [[13]](https://www.audiolabs-erlangen.de/resources/MIR/FMP/C4/C4S4_NoveltySegmentation.html) [[12]](https://biblio.ugent.be/publication/8577611)

Recurrence asks whether a passage sounded like something earlier. McFee and Ellis combined repeating material with local continuity and used spectral clustering to recover groups at several levels. This can connect two choruses separated by a verse. A local novelty peak alone supplies little evidence that the new passage is a returning chorus. MSAF provides a common interface for comparing structural algorithms on recordings. [[14]](https://brianmcfee.net/papers/ismir2014_spectral.pdf) [[15]](https://github.com/urinieto/msaf)

### Learn the intended level

Jiang and colleagues modeled nested metrical groupings. Their earlier work used symbolic scores. The later implementation accepts audio and needs a reliable beat grid. For conditional random field (CRF) decoding of the larger group hierarchy it also needs accurate downbeats. An erroneous bar grid therefore propagates into the larger group prediction. [[16]](https://github.com/music-x-lab/Self-Supervised-Metrical-Structure)

All-In-One predicts functional segments such as verse and chorus along with beat and downbeat positions. For explicitly annotated local phrases, the **FaiSonic MMD preprint** supplies a different target. Its 500 Chinese- and English-language pop tracks have separate phrase and section references. Its joint model reports phrase F1 of **62.90% within ±0.5 seconds** and **79.18% within ±3 seconds**. That larger window spans over six beats at 124 BPM. These are offline results for that pop corpus. [[19]](https://arxiv.org/abs/2307.16425) [[18]](https://arxiv.org/abs/2611.13503)

### Give the performer three distinct controls

Show the **predicted group start** as a countdown. Show **observed arrangement changes** as evidence with a time and confidence. Make **manual phrase start** an immediate authoritative anchor. For an automatic hard cut, require agreement between the bar phase and the selected cue type, or allow the operator to arm the next downbeat. While confidence is weak, gradual palette or density changes cost less than an unwanted scene cut. This is a proposed performance policy. Beat Link Trigger illustrates another approach on compatible setups, using prepared rekordbox phrase information to place performance cues. [[34]](https://blt-guide.deepsymmetry.org/beat-link-trigger/main/Shows_PhraseTriggers.html) [[17]](https://mir-aidj.github.io/raveform/)

## Neural rhythm tracking

Compare recurrent networks, particle filters, transformers, and phrase-specific models.

A neural beat tracker learns which combinations of timbre, rhythm, and musical context predict a labeled event. Its output might be one tempo class per excerpt, per-frame beat and downbeat activations, or named functional sections. The decoder that turns these outputs into events can change the behavior as much as the network itself. [[22]](https://ismir2018.ircam.fr/doc/pdfs/141_Paper.pdf) [[21]](https://www.cp.jku.at/research/papers/Boeck_etal_ISMIR_2016.pdf) [[26]](https://ismir2024program.ismir.net/poster_10.html) [[19]](https://arxiv.org/abs/2307.16425)

### From features to decisions

A typical front end maps a waveform to a time–frequency representation, often a log-mel spectrogram. A convolutional network learns local patterns across time and frequency. A recurrent network carries context from frame to frame. A bidirectional network can use sounds on both sides of the current frame. Transformers learn which other frames or frequency bands are useful, with their effective context set by masks, windows, and architecture. [[21]](https://www.cp.jku.at/research/papers/Boeck_etal_ISMIR_2016.pdf) [[25]](https://arxiv.org/abs/2209.07140) [[26]](https://ismir2024program.ismir.net/poster_10.html)

A model trained on frame-level beat labels produces an activation curve. Peak picking is inexpensive but can admit extra events or uneven spacing. A DBN adds constraints on tempo and meter. Particle filtering is another option for updating several rhythmic hypotheses as the audio arrives. Learned anticipation shifts the training targets so an event can be predicted early enough for a bounded peak picker. Throughput faster than playback tells you little about future context or decision latency. [[21]](https://www.cp.jku.at/research/papers/Boeck_etal_ISMIR_2016.pdf) [[23]](https://arxiv.org/abs/2108.03576) [[27]](https://ismir2026program.ismir.net/poster_187.html)

### What the main model families return

| Model | What it learns and returns | Where it fits |
|---|---|---|
| Tempo-CNN | A tempo class from an 11.9 s mel excerpt. Sliding it yields a tempo curve. | A BPM estimate or candidate prior. Beat times require another stage. [[22]](https://ismir2018.ircam.fr/doc/pdfs/141_Paper.pdf) |
| madmom joint tracker | Bidirectional recurrent beat/downbeat activations followed by a meter-aware DBN. | An offline reference grid. The classic joint path uses future context. [[21]](https://www.cp.jku.at/research/papers/Boeck_etal_ISMIR_2016.pdf) |
| BeatNet / BeatNet+ | Causal convolutional/recurrent evidence with sequential inference. BeatNet uses two particle filters and can infer meter. BeatNet+ adds training on drum-reduced material. | First candidates to test on a stream, including a breakdown where percussion recedes. The published implementation documents microphone input. [[23]](https://arxiv.org/abs/2108.03576) [[24]](https://transactions.ismir.net/articles/10.5334/tismir.198) |
| Beat Transformer | Dilated self-attention across time and demixed instrument channels. | File analysis where separated bass, drums, and harmonic parts can clarify accents. Budget for separation and context. [[25]](https://arxiv.org/abs/2209.07140) |
| Beat This! | Convolutions and frequency/time attention, a timing-tolerant loss, and direct event extraction without a mandatory DBN. | Offline file grids. The authors report stronger F1 alongside weaker continuity on some measures. [[26]](https://ismir2024program.ismir.net/poster_10.html) |
| Beat This! Live | A causal adaptation with bounded temporal context and targets shifted to anticipate events. | A streaming research candidate. Its repository emphasizes training and evaluation. Measure end-to-end latency in the target runtime. [[27]](https://ismir2026program.ismir.net/poster_187.html) |
| All-In-One | Joint beat/downbeat activations and functional boundaries and labels from separated spectrograms with neighborhood attention. | Pre-analysis of an arrangement. A chorus label does not enumerate the melodic phrases inside it. [[19]](https://arxiv.org/abs/2307.16425) [[18]](https://arxiv.org/abs/2611.13503) |

### A comparison with a stated protocol

On **GTZAN**, the Beat This! Live authors report online beat/downbeat F1 of **86.7 / 74.1**, compared with **80.6 / 56.5** for BeatNet+ in their comparison. Downbeats are the harder target on this test. The result motivates a side-by-side trial on live mixes, especially one that records cold-start lock, stable phase through a breakdown, and the delay between prediction and a displayed frame. These are published scores, independent of this guide’s synthetic teaching loop. [[27]](https://ismir2026program.ismir.net/poster_187.html)

### Open code, weights, and deployment

The Beat This! and Beat This! Live repositories specify MIT for code and released weights. madmom’s repository gives its code a BSD license and distributes its model/data files under CC BY-NC-SA 4.0 by default. BTrack’s C++ repository uses GPL-3.0. These distinctions matter when choosing components to redistribute with a visualizer. For a web prototype, a local native or Python analysis service can expose timestamped events to the renderer. Revalidate output parity and timing before swapping in a browser model or quantized weights. [[26]](https://ismir2024program.ismir.net/poster_10.html) [[27]](https://ismir2026program.ismir.net/poster_187.html) [[21]](https://www.cp.jku.at/research/papers/Boeck_etal_ISMIR_2016.pdf) [[8]](https://github.com/adamstark/BTrack)

## Driving a visualizer

Connect audio features and transport clocks to a renderer with predictable timing.

For a 4/4 VJ rig, start with the clock the source can actually provide. A prepared file allows whole-song analysis and human corrections. A live DJ transport can share timing even during silence. An unknown stereo feed has only the samples received so far. The renderer can consume the same event format for each source while preserving how far ahead it can trust each event. [[32]](https://ableton.github.io/link/) [[9]](https://www.audiolabs-erlangen.com/resources/MIR/2024-TISMIR-RealTimePLP) [[19]](https://arxiv.org/abs/2307.16425)

### Choose the path that matches the input

| Input | Suggested analysis path | Operator control |
|---|---|---|
| Music on Windows, including system playback | Capture the selected output with WASAPI loopback in a native companion, or use browser display capture when the selected surface offers audio. Check that an audio track was returned. Compare real-time PLP and BeatNet on the same feed. | Tap tempo, half/double, set “one,” and mark phrase start. [[35]](https://developer.mozilla.org/en-US/docs/Web/API/MediaDevices/getDisplayMedia) [[9]](https://www.audiolabs-erlangen.com/resources/MIR/2024-TISMIR-RealTimePLP) [[23]](https://arxiv.org/abs/2108.03576) |
| Traktor with Link | Use Link for tempo and phase, with a compatible bridge for a browser renderer. Observe drift and align the deck’s beatgrid. Use the audio path for timbre, attacks, and confirmation. | Reset the downbeat independently and arm transitions. [[32]](https://ableton.github.io/link/) [[33]](https://support.native-instruments.com/support/solutions/articles/69000879473-how-to-sync-traktor-to-other-applications-via-ableton-link) |
| Loaded audio file | Precompute beats/downbeats with Beat This!, compare the grid with a classical tracker, and propose sections with All-In-One. Save corrected cues keyed to the exact audio version. | Scrub to the drop, move a marker, or accept and label a phrase. [[26]](https://ismir2024program.ismir.net/poster_10.html) [[7]](https://librosa.org/doc/main/api/generated/librosa.beat.beat_track.html) [[19]](https://arxiv.org/abs/2307.16425) |
| Compatible prepared DJ hardware | Read existing track and phrase data where available. Beat Link Trigger illustrates using that information for start, body, fill, and end cues. | Give hand-authored track cues priority. [[34]](https://blt-guide.deepsymmetry.org/beat-link-trigger/main/Shows_PhraseTriggers.html) |

Essentia.js exposes audio algorithms through WebAssembly. Inspect each algorithm’s input requirements. Essentia’s RhythmExtractor2013 uses full-track statistics and is recommended for offline work in its own tutorial. A streaming API that accepts chunks does not by itself make the resulting beat decisions causal. For lightweight in-browser continuous features, Meyda is a useful starting point. [[28]](https://essentia.upf.edu/tutorial_rhythm_beatdetection.html) [[29]](https://meyda.js.org/audio-features)

### Map each signal to the right visual timescale

| Audio or clock feature | A concrete visual mapping | Behavior to test |
|---|---|---|
| Smoothed RMS and band energies | Bass → shape size, midrange → layer density, treble → fine detail. Tune attack and release separately. | Loud masters, quiet breakdowns, and gain steps should remain legible. [[29]](https://meyda.js.org/audio-features) [[30]](https://github.com/projectM-visualizer/projectm/wiki/Preset-Authoring-Guide) |
| Spectral flux and centroid | Use transients to accent line weight. Use spectral brightness for a slow change in texture. | A filter sweep can brighten continuously without extra beat triggers. [[1]](https://cs-web.bu.edu/fac/snyder/cs583/Literature%20and%20Resources/bello_onset_tutorial.pdf) [[29]](https://meyda.js.org/audio-features) |
| Continuous beat phase | Derive a periodic motion curve and put its apex at the predicted beat. | Motion should cross a brief missing kick smoothly. [[9]](https://www.audiolabs-erlangen.com/resources/MIR/2024-TISMIR-RealTimePLP) |
| Downbeat and group position | Rotate a palette per confirmed bar or morph a shape across eight manually anchored bars. | A phase correction should avoid duplicate cues. [[32]](https://ableton.github.io/link/) [[11]](https://www.eurasip.org/Proceedings/Eusipco/Eusipco2006/papers/1568981125.pdf) |
| Predicted or confirmed section cue | Arm a scene change, then commit on a chosen downbeat. Use observed changes for effects that can arrive later. | Count false hard cuts during fills, crossfades, and long drops. [[31]](https://personales.upv.es/jjlopez/publications/SegmentEDM_AES144th.pdf) [[17]](https://mir-aidj.github.io/raveform/) |

projectM’s MilkDrop-compatible preset system exposes quick and damped bass, midrange, and treble responses. Butterchurn is a browser implementation in this family. Both are useful references for the visual side of a system even if a separate component supplies the musical clock. [[30]](https://github.com/projectM-visualizer/projectm/wiki/Preset-Authoring-Guide)

### Schedule on the audio timeline

Timestamp analysis frames in audio sample time. A useful event includes its musical time, its emission time, event kind, confidence or stability, and the transport generation that produced it. For a known file, use the playback position and rate to map cue times onto presentation time. On a seek or loop, increment the generation and cancel pending cues from the old one. A causal system can extrapolate the next beat and revise future predictions as evidence arrives. [[36]](https://developer.mozilla.org/en-US/docs/Web/API/AudioContext/getOutputTimestamp) [[9]](https://www.audiolabs-erlangen.com/resources/MIR/2024-TISMIR-RealTimePLP)

In the browser, keep audio-thread work short. An AudioWorklet can gather timestamped feature frames while a worker or native process runs heavier inference. Schedule visual updates with the display’s animation callback and evaluate phase for the expected presentation time. Web Audio’s getOutputTimestamp pairs an output-context time with performance time. It does not measure input capture, a projector’s processing, or the path from a separate speaker to the audience. Measure and store a route-specific offset. [[36]](https://developer.mozilla.org/en-US/docs/Web/API/AudioContext/getOutputTimestamp)

For example, at 60 Hz an entire display frame takes about 16.7 ms. Measure the signed difference between an audible transient and a co-recorded on-screen flash. Use a shared recorder and account for its own audio/video offset, or use a light sensor plus audio loopback for finer measurements. Correct fixed delay separately from jitter. Confidence can fall independently for the beat, bar, or phrase, so a doubtful phrase prediction need not stop the beat animation.

## Measuring reliability

Test timing, continuity, startup, and false scene changes on the right annotations.

“Reliable” needs a target and a deadline. A BPM classifier, a pulse tracker, and a system that triggers the first bar of a drop can all be scored on the same song and give very different answers. Report musical accuracy and the delay at which each answer was available. [[37]](https://transactions.ismir.net/articles/10.5334/tismir.43) [[38]](https://mir-eval.readthedocs.io/latest/api/beat.html) [[27]](https://ismir2026program.ismir.net/poster_187.html)

### Match the reference to the question

| Reference | What it can test | What requires additional labels |
|---|---|---|
| GiantSteps Tempo | Tempo on 664 electronic music previews, including revised reference values. | Beat placement, downbeats, and phrase starts. [[39]](https://github.com/GiantSteps/giantsteps-tempo-dataset) |
| Harmonix Set | Beats, metrical positions, and functional sections on 912 popular tracks. | Every internal phrase of a section or the precise cue a VJ would choose. [[40]](https://github.com/urinieto/harmonixset) |
| Raveform | EDM structure including drops and breakdowns, with 1,423 structure-annotated individual tracks and beat/downbeat information. | An arbitrary DJ’s live transition or every phrase in the mix. [[17]](https://mir-aidj.github.io/raveform/) |
| FaiSonic MMD | Explicitly separate phrase and section boundaries in a 500-song pop corpus. | Generalization to live EDM or the performer’s preferred mixing points. [[18]](https://arxiv.org/abs/2611.13503) |
| A locally annotated set recording | Lock and recovery across actual crossfades, transport jumps, silent passages, and the display chain. | Reproducible conclusions until the annotation policy, hardware, and sample are described. |

### Use metrics that expose the failure

**Tempo.** Report absolute BPM error and accuracy at a stated relative tolerance. A common Accuracy 1 convention uses 4%. Accuracy 2 also accepts metrical alternatives such as double, half, triple, and one third. Publish both and count octave errors separately. A display reading 62 on a 124 BPM track has very different consequences for motion depending on which level the performer wanted. [[37]](https://transactions.ismir.net/articles/10.5334/tismir.43)

**Beats and downbeats.** Match event times one-to-one against their own reference lists. Report precision, recall, and F1 at a stated window. The common mir_eval beat default is ±70 ms. A system that is consistently 60 ms late can still match every beat. Report signed bias, median absolute error, and the 95th percentile absolute error for matched events, along with misses. CMLc measures the longest uninterrupted correct-level run. CMLt credits all correct-level spans. AML variants allow alternative metrical interpretations. Measure downbeat F1 separately. [[38]](https://mir-eval.readthedocs.io/latest/api/beat.html)

**Phrases and sections.** Use matching at a tight tolerance suited to the cue and also report conventional structural windows such as ±0.5 s and ±3 s for comparison. Score the nominated downbeat or bar index for DJ grouping. Score a melodic pickup at its actual timestamp when that is the target. For labeled sections, evaluate labels as well as boundaries. Record false hard cuts per hour and how often predictions arrive before their cue deadlines. [[38]](https://mir-eval.readthedocs.io/latest/api/beat.html) [[18]](https://arxiv.org/abs/2611.13503) [[17]](https://mir-aidj.github.io/raveform/)

### A repeatable live test

1. Choose held-out tracks and mixes. Keep tracks from the same recording, remix family, or performance out of both tuning and test sets. Document meter, tempo range, style, and what counts as a phrase.
2. Have two listeners mark beats, downbeats, and the specific VJ cues. Retain disagreements and accept multiple defensible starts where appropriate. Distinguish measured annotations from a grid imported from software under test.
3. Feed the same PCM to each candidate in timestamped chunks at playback speed. Log the audio timestamp, wall-clock decision time, intended visual time, and frames actually presented. For prepared-file runs, label whole-track access explicitly.
4. Start the stream mid-bar, remove the kick, change tempo, introduce a two-bar extension, crossfade two decks, then seek. Record cold-start time, longest wrong-phase run, reacquisition time, false doubles, and missed cues. Common beat evaluations trim the first five seconds. Keep an untrimmed startup result. [[38]](https://mir-eval.readthedocs.io/latest/api/beat.html)
5. Compare manual anchoring, a classical baseline, and a causal neural tracker on the same material. Report per-track results and worst cases as well as averages. Log CPU time and missed render deadlines on the machine driving the projector.

For a first klobvision comparison, use real-time PLP and BeatNet on an identical live feed, Beat This! on the corresponding files, and a manually checked cue sheet for phrase starts. Add Beat This! Live as a third online candidate. Compare how long each stays aligned through the first drumless break and how many scene changes the performer would have reversed. This is an evaluation design. The published model comparisons and their datasets are identified in the neural chapter. [[9]](https://www.audiolabs-erlangen.com/resources/MIR/2024-TISMIR-RealTimePLP) [[23]](https://arxiv.org/abs/2108.03576) [[26]](https://ismir2024program.ismir.net/poster_10.html) [[27]](https://ismir2026program.ismir.net/poster_187.html)

## Method index

| Method | Output | Timing | How it works | Tradeoff |
|---|---|---|---|---|
| Envelope threshold | Onsets | Live | Compare short-window energy with a recent baseline, then debounce attacks. | Very low cost. A sustained pad or a snare roll can swamp a simple threshold. [[1]](https://cs-web.bu.edu/fac/snyder/cs583/Literature%20and%20Resources/bello_onset_tutorial.pdf) |
| Spectral flux / SuperFlux | Onsets | Live-capable | Measure positive changes across frequency bands. SuperFlux suppresses motion caused by vibrato. | Effective on attacks. Peak picking that waits for future frames adds delay. [[2]](https://www.dafx.de/paper-archive/details/0oee-99Z88WL7pSo749gcA) |
| Inter-onset interval voting | Tempo | Live-capable | Cluster gaps between onsets and vote for a plausible pulse period. | Cheap and interpretable. Subdivisions, syncopation, and missing notes split the votes. [[5]](https://webspace.eecs.qmul.ac.uk/s.e.dixon/pub/2007/jnmr07.pdf) |
| Autocorrelation / tempogram | Tempo | Live-capable | Score periodicity over many lags and update the scores in local windows. | Reveals multiple tempo candidates. Longer history improves resolution and slows adaptation. [[4]](https://www.audiolabs-erlangen.de/resources/MIR/FMP/C6/C6S2_TempogramAutocorrelation.html) |
| Comb filter bank | Tempo | Live-capable | Compare resonator responses at candidate beat periods across frequency bands. | Works on streams. Ringing and build-up affect response to a tempo change. [[3]](https://pubmed.ncbi.nlm.nih.gov/9440344/) |
| Grid fit / phase-locked loop (PLL) | Beats | Live-capable | Align an oscillator to nearby onset evidence and smooth phase corrections. | Low compute and predictable motion. Can lock halfway between true beats. [[12]](https://biblio.ugent.be/publication/8577611) |
| BeatRoot agents | Beats | Sequence model | Competing agents follow candidate periods and onset alignments. | Can recover from a false start. Published system and live adaptations need separate latency tests. [[5]](https://webspace.eecs.qmul.ac.uk/s.e.dixon/pub/2007/jnmr07.pdf) |
| Dynamic programming | Beats | Offline | Optimize onset evidence and interval consistency over the recording. | Bridges weak evidence. Backtracking uses future material. [[6]](https://www.ee.columbia.edu/~dpwe/pubs/Ellis07-beattrack.pdf) |
| BTrack / real-time PLP | Beats | Live | Track or project a local pulse. PLP supplies stability and future beat candidates. | A good interpretable baseline. Sudden stops can invalidate projected events. [[9]](https://www.audiolabs-erlangen.com/resources/MIR/2024-TISMIR-RealTimePLP) |
| Bar-phase templates | Downbeats | Live-capable | Score metrical offsets using bass, timbre, and repeated accents on a known beat grid. | Simple and inspectable. Uniform kicks leave several offsets equally plausible. [[11]](https://www.eurasip.org/Proceedings/Eusipco/Eusipco2006/papers/1568981125.pdf) |
| Bar counter + manual anchor | Phrases | Live | Count bar groups from an operator-provided musical start. | Predicts ahead. A two-bar extension displaces later cues until re-anchored. [[17]](https://mir-aidj.github.io/raveform/) |
| Checkerboard novelty | Phrases | Offline | Look for changes between neighborhoods in a self-similarity matrix. | Finds acoustic boundaries. A symmetric window spends future context. [[13]](https://www.audiolabs-erlangen.de/resources/MIR/FMP/C4/C4S4_NoveltySegmentation.html) |
| Recurrence / spectral clustering | Sections | Offline | Combine repeated musical content with local continuity in a graph. | Links returning passages. Cluster resolution changes the structural level. [[14]](https://brianmcfee.net/papers/ismir2014_spectral.pdf) |
| Tempo-CNN | Tempo | Windowed | Classify local tempo from an 11.9-second mel-spectrogram excerpt. | Gives rate without a beat offset. Requires a substantial observation window. [[22]](https://ismir2018.ircam.fr/doc/pdfs/141_Paper.pdf) |
| RNN + DBN / madmom | Downbeats | Offline | Learn joint beat/downbeat activations, then decode a plausible meter-aware path. | Enforces rhythmic consistency. The classic joint model uses bidirectional evidence. [[21]](https://www.cp.jku.at/research/papers/Boeck_etal_ISMIR_2016.pdf) |
| BeatNet / BeatNet+ | Downbeats | Live | Extract causal neural activations. BeatNet uses particle filters for beat phase and bar state. | Supports live tracking. Test recovery on percussion-free passages and blends. [[23]](https://arxiv.org/abs/2108.03576) |
| Beat This! | Downbeats | Offline | Use convolutions and attention for beat/downbeat activations with direct decoding. | Strong published F1. Grid continuity can still suffer. [[26]](https://ismir2024program.ismir.net/poster_10.html) |
| Beat This! Live | Downbeats | Live | Use causal context and shifted targets to anticipate beat/downbeat decisions. | Promising published online scores. Profile full integration and scheduling. [[27]](https://ismir2026program.ismir.net/poster_187.html) |
| All-In-One | Sections | Offline | Jointly infer beats, downbeats, functional section boundaries, and labels from stems. | Useful for precomputed scene cues. Labels target musical function. [[19]](https://arxiv.org/abs/2307.16425) |
| Learned phrase boundaries | Phrases | Offline | Train on independent phrase annotations and decode candidate boundaries. | Closer to melodic phrasing. Scores depend strongly on corpus and matching tolerance. [[18]](https://arxiv.org/abs/2611.13503) |

## Annotated sources

1. **[A tutorial on onset detection in music signals](https://cs-web.bu.edu/fac/snyder/cs583/Literature%20and%20Resources/bello_onset_tutorial.pdf)**. Bello et al. · IEEE, 2005. What changes in energy, spectrum, or phase when a note begins.

2. **[Maximum filter vibrato suppression for onset detection](https://www.dafx.de/paper-archive/details/0oee-99Z88WL7pSo749gcA)**. Böck & Widmer · DAFx, 2013. SuperFlux reduces false attacks caused by vibrato using a frequency-axis maximum filter.

3. **[Tempo and beat analysis of acoustic musical signals](https://pubmed.ncbi.nlm.nih.gov/9440344/)**. Scheirer · JASA, 1998. A bank of tuned comb filters extracts periodicity in subband envelopes.

4. **[Fourier and autocorrelation tempograms](https://www.audiolabs-erlangen.de/resources/MIR/FMP/C6/C6S2_TempogramAutocorrelation.html)**. Müller · Fundamentals of Music Processing notebooks. Worked examples of two different ways to measure local periodicity, including their harmonic ambiguities. [Fourier notebook](https://www.audiolabs-erlangen.de/resources/MIR/FMP/C6/C6S2_TempogramFourier.html).

5. **[Evaluation of the audio beat tracking system BeatRoot](https://webspace.eecs.qmul.ac.uk/s.e.dixon/pub/2007/jnmr07.pdf)**. Dixon · Journal of New Music Research, 2007. Clusters inter-onset intervals and lets competing tempo and phase hypotheses follow the signal.

6. **[Beat tracking by dynamic programming](https://www.ee.columbia.edu/~dpwe/pubs/Ellis07-beattrack.pdf)**. Ellis · Journal of New Music Research, 2007. Balances evidence for each beat with a penalty for implausible spacing. [Author’s example](https://www.ee.columbia.edu/~dpwe/LabROSA/matlab/beat_simple/).

7. **[librosa beat tracker](https://librosa.org/doc/main/api/generated/librosa.beat.beat_track.html)**. librosa contributors · Python. An accessible implementation of the onset, tempo, and dynamic-programming pipeline. [Repository](https://github.com/librosa/librosa).

8. **[BTrack](https://github.com/adamstark/BTrack)**. Stark, Davies & Plumbley · DAFx, 2009. A C++ beat tracker designed for real-time, beat-synchronous effects. [Paper](https://www.adamstark.co.uk/pdf/papers/beat-synchronous-DAFX-2009.pdf).

9. **[A Real-Time Beat Tracking System with Zero Latency and Enhanced Controllability](https://www.audiolabs-erlangen.com/resources/MIR/2024-TISMIR-RealTimePLP)**. Meier, Chiu & Müller · TISMIR, 2024. A causal predominant local pulse implementation with beat stability and forward projections. [Code and live demos](https://github.com/groupmm/real_time_plp).

10. **[aubio tempo API](https://aubio.org/doc/latest/tempo_8h.html)**. aubio contributors · C. A streaming tempo object exposes the latest beat, period, BPM, and confidence. [Repository](https://github.com/aubio/aubio).

11. **[A spectral difference approach to downbeat extraction in musical audio](https://www.eurasip.org/Proceedings/Eusipco/Eusipco2006/papers/1568981125.pdf)**. Davies & Plumbley · EUSIPCO, 2006. Tests alternative bar offsets with beat-synchronous spectral evidence and a known meter.

12. **[From raw audio to a seamless mix](https://biblio.ugent.be/publication/8577611)**. Vande Veire & De Bie · EURASIP, 2018. An evaluated drum-and-bass auto-DJ with separate beat, downbeat, segmentation, and transition stages. [Publisher’s full text](https://link.springer.com/article/10.1186/s13636-018-0134-8).

13. **[Novelty-based segmentation](https://www.audiolabs-erlangen.de/resources/MIR/FMP/C4/C4S4_NoveltySegmentation.html)**. Müller · Fundamentals of Music Processing notebooks. An executable treatment of Foote’s checkerboard kernel, self-similarity, and window scale. [Foote’s original paper](https://doi.org/10.1109/ICME.2000.869637).

14. **[Analyzing song structure with spectral clustering](https://brianmcfee.net/papers/ismir2014_spectral.pdf)**. McFee & Ellis · ISMIR, 2014. A recurrence graph and local acoustic continuity expose structural repetition at several scales.

15. **[Music Structure Analysis Framework](https://github.com/urinieto/msaf)**. Nieto et al. · Python. A common interface for running and evaluating structural segmentation algorithms.

16. **[Deep self-supervised hierarchical metrical structure modeling](https://github.com/music-x-lab/Self-Supervised-Metrical-Structure)**. Jiang et al. · ICASSP, 2023. The authors’ audio implementation documents required beat and downbeat annotations for hierarchical decoding. [Earlier symbolic model](https://archives.ismir.net/ismir2022/paper/000023.pdf).

17. **[Raveform](https://mir-aidj.github.io/raveform/)**. Kim et al. · TISMIR, 2026. EDM beat and downbeat information and 1,423 tracks with functional structure annotations. Includes drop and breakdown examples. [Paper](https://reference-global.com/download/article/10.5334/tismir.288.pdf).

18. **[FaiSonic MMD](https://arxiv.org/abs/2611.13503)**. Zhang et al. · Preprint, November 2026. Separate phrase and section annotations and multitask baselines for 500 Chinese- and English-language pop tracks.

19. **[All-In-One metrical and functional structure analysis](https://arxiv.org/abs/2307.16425)**. Kim & Nam · WASPAA, 2023. A joint model of beats, downbeats, functional boundaries, and labels using separated stems. [Repository](https://github.com/mir-aidj/all-in-one).

20. **[Cue point estimation using object detection](https://arxiv.org/abs/2407.06823)**. Argüello, Lanzendörfer & Wattenhofer · 2024. Learns candidate DJ transition points from human cue annotations. The annotated target differs from a phrase boundary.

21. **[Joint beat and downbeat tracking with recurrent neural networks](https://www.cp.jku.at/research/papers/Boeck_etal_ISMIR_2016.pdf)**. Böck, Krebs & Widmer · ISMIR, 2016. Bidirectional recurrent evidence and a probabilistic meter-aware decoder. The implementation has separate code and model licenses. [Code and model licenses](https://github.com/CPJKU/madmom/blob/main/LICENSE).

22. **[A single-step approach to musical tempo estimation using a convolutional neural network](https://ismir2018.ircam.fr/doc/pdfs/141_Paper.pdf)**. Schreiber & Müller · ISMIR, 2018. Tempo classification from a short mel-spectrogram excerpt, with a local-tempo visualization. [Repository](https://github.com/hendriks73/tempo-cnn).

23. **[BeatNet](https://arxiv.org/abs/2108.03576)**. Heydari, Cwitkowitz & Duan · ISMIR, 2021. Causal CRNN plus two particle filters to infer beats, downbeats, and meter online. [Streaming code](https://github.com/mjhydri/BeatNet).

24. **[BeatNet+](https://transactions.ismir.net/articles/10.5334/tismir.198)**. Heydari et al. · TISMIR, 2024. Dual-branch training includes musical evidence from material with reduced percussion. [Repository](https://github.com/mjhydri/BeatNet-Plus).

25. **[Beat Transformer](https://arxiv.org/abs/2209.07140)**. Zhao et al. · ISMIR, 2022. Demixed instrument channels and dilated self-attention for beat and downbeat tracking.

26. **[Beat This! Accurate beat tracking without DBN postprocessing](https://ismir2024program.ismir.net/poster_10.html)**. Foscarin, Schlüter & Widmer · ISMIR, 2024. Offline beat and downbeat inference, time-tolerant training, and a useful continuity tradeoff. [Code and weights](https://github.com/CPJKU/beat_this).

27. **[Beat This! Live](https://ismir2026program.ismir.net/poster_187.html)**. Juhász, Widmer & Schlüter · ISMIR, 2026. Causal, streaming-oriented adaptation with anticipatory targets and a controlled online GTZAN comparison. [Code and weights](https://github.com/CPJKU/beat_this_live).

28. **[Essentia rhythm extraction and Essentia.js](https://essentia.upf.edu/tutorial_rhythm_beatdetection.html)**. Music Technology Group, Universitat Pompeu Fabra. Whole-track RhythmExtractor2013 and the WebAssembly port for browser feature extraction. [Essentia.js](https://mtg.github.io/essentia.js/).

29. **[Meyda audio features](https://meyda.js.org/audio-features)**. Meyda contributors · JavaScript. Feature definitions for RMS, spectral centroid, chroma, and other visual control signals. [Repository](https://github.com/meyda/meyda).

30. **[projectM preset authoring guide](https://github.com/projectM-visualizer/projectm/wiki/Preset-Authoring-Guide)**. projectM contributors · MilkDrop-compatible visualizer. Immediate and damped band energies as inputs to animations and shaders. [Browser implementation · Butterchurn](https://github.com/jberg/butterchurn).

31. **[Low cost algorithm for online segmentation of Electronic Dance Music](https://personales.upv.es/jjlopez/publications/SegmentEDM_AES144th.pdf)**. Aguilera et al. · AES, 2018. A live system detects breaks and changes in instrumentation from frequency-band energy.

32. **[Ableton Link documentation](https://ableton.github.io/link/)**. Ableton · Synchronization protocol. The meaning of beat alignment, tempo, phase, and quantum in linked applications. [Repository](https://github.com/Ableton/link).

33. **[Syncing Traktor to other applications via Ableton Link](https://support.native-instruments.com/support/solutions/articles/69000879473-how-to-sync-traktor-to-other-applications-via-ableton-link)**. Native Instruments · Traktor. Documents a phase-offset case and the Reset Downbeat controller mapping.

34. **[Beat Link Trigger phrase triggers](https://blt-guide.deepsymmetry.org/beat-link-trigger/main/Shows_PhraseTriggers.html)**. Deep Symmetry · Performance software. Drives visual and lighting cues from prepared rekordbox phrase analysis on compatible setups. [Repository](https://github.com/Deep-Symmetry/beat-link-trigger).

35. **[System audio capture on the web and Windows](https://developer.mozilla.org/en-US/docs/Web/API/MediaDevices/getDisplayMedia)**. MDN & Microsoft · Platform documentation. Browser display capture has optional audio. Native WASAPI loopback captures the system mix on a render endpoint. [WASAPI loopback](https://learn.microsoft.com/en-us/windows/win32/coreaudio/loopback-recording).

36. **[Web Audio timing and AudioWorklet](https://developer.mozilla.org/en-US/docs/Web/API/AudioContext/getOutputTimestamp)**. MDN · Web Audio API. Relates the audio output clock to performance time and documents the audio rendering thread. [AudioWorklet](https://developer.mozilla.org/en-US/docs/Web/API/AudioWorklet).

37. **[Music tempo estimation: Are we done yet?](https://transactions.ismir.net/articles/10.5334/tismir.43)**. Schreiber, Urbano & Müller · TISMIR, 2020. Explains metrical ambiguity and the effects of benchmark design on tempo claims.

38. **[mir_eval beat and segmentation metrics](https://mir-eval.readthedocs.io/latest/api/beat.html)**. Raffel et al. · Evaluation library. Documents matching windows, continuity at different metrical levels, trimming, and boundary metrics. [Segment metrics](https://mir-eval.readthedocs.io/latest/api/segment.html).

39. **[GiantSteps Tempo](https://github.com/GiantSteps/giantsteps-tempo-dataset)**. Knees et al. · ISMIR, 2015. Tempo and genre annotations for 664 electronic music previews, including revised tempo references.

40. **[The Harmonix Set](https://github.com/urinieto/harmonixset)**. Nieto et al. · ISMIR, 2019. 912 popular tracks with beats, positions within the bar, and functional sections.
