AudioLabs - Automated Real-Time Beat Tracking: Response Time and Confidence Analysis

Automated Real-Time Beat Tracking: Response Time and Confidence Analysis

This is the accompanying website for the Bachelor Thesis [1].

Rico Rosenbusch
Automated Real-Time Beat Tracking: Response Time and Confidence Analysis
Bachelor Thesis, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), 2024. Demo

@misc{Rosenbusch24_ARTBeaT_Bachelor_Thesis,
title       = {Automated Real-Time Beat Tracking: Response Time and Confidence Analysis},
author      = {Rico Rosenbusch},
year        = {2024},
note        = {Bachelor Thesis, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU)},
url-demo    = {https://audiolabs-erlangen.de/resources/MIR/2024-ARTBeaT}
}

Abstract

Real-time beat tracking remains a challenging task for music information retrieval. This thesis presents a comparative analysis of human and machine real-time beat tracking performance. The algorithm used is based on the predominant local pulse (PLP) procedure. For comparing the results of the algorithm and humans with each other, their performances are evaluated on a dataset of 25 short unfamiliar pieces of music, containing rhythmic challenges such as tempo changes, metric ambiguity, or syncopated rhythms. A web-based listening test was conducted to collect real-time annotations from participants tapping along to each example. The metrics used to evaluate the examples annotated by the algorithm and users include F-measure, response time, and confidence measurements. Note that the algorithm is not compared to individuals but to the average performance of all users, obtained by using a kernel density estimation (KDE) technique. Parameter tuning is also explored, showing the effect of kernel size on the algorithm's stability and responsiveness. Differences in methodology, strengths, and limitations are identified between human and machine rhythmic cognition. The analysis provides new insights into more natural modeling of human timing perception. Humans show faster adaptation after tempo changes but are also briefly confused by deceptive rhythms. With a well-tuned set of parameters, the algorithm can match or exceed human consistency on some rhythms. Overall, this work advances understanding of real-time beat tracking through direct comparison with human annotation data. The findings have implications for improving computational rhythmic analysis and further outlining similarities and divergences between human and machine rhythmic cognition.

Analysis Example

The figure presents the analysis of a single audio example from the dataset, highlighting both the algorithm's beat tracking output and the human user annotations collected through a listening test.

Figure 4.21: Visualization of important audio, online PLP and user annotation data of audio example 13_ARTBeaT_TempoA_180to120.wav. (a) Waveform. (b) Spectrogram. (c) Novelty function. (d) Peaks detected by online PLP algorithm. (e) β-confidence. (f) Histogram of user annotations, with a Gaussian KDE and its peaks plotted over it. (g) user confidence. (h) user recall ratio (URR).

Dataset

We present the dataset "ARTBeaT," an acronym for the title of this bachelor thesis, "Automated Real-Time Beat Tracking."

Download:

ARTBeaT Dataset by Rico Rosenbusch is licensed under CC BY 4.0

Files

The files are named using the following pattern: {ID}_ARTBeaT_{Category}_{Name}.wav, as shown in the table below.

ID	Category	Name	Description	Length
01	basic	calib	Short metronome for calibration.	9.7 s
02	basic	arp1	Starts with an arpeggio played on a synthesizer, continues in same tempo and measure with added instruments.	17.5 s
03	rhythmic deception	arp2	Starts with same arpeggio, but the other instruments come in 3 8th notes later.	18.1 s
04	polyrhythm	arp3	Starts with the same arpeggio, continues with different drumbeat at 2/3 the tempo playing simultaneously with original arpeggio.	23.1 s
05	abrupt tempo change	75to150	Simple drumbeat abruptly changing tempo from 75 to 150 BPM.	13.6 s
06	abrupt tempo change	150to75	Simple drumbeat abruptly changing tempo from 150 to 75 BPM.	13.6 s
07	abrupt tempo change	75to112.5	Simple drumbeat abruptly changing tempo from 75 to 112.5 BPM.	13.3 s
08	abrupt tempo change	112.5to75	Simple drumbeat abruptly changing tempo from 112.5 to 75 BPM.	14.0 s
09	abrupt tempo change	90to80	Simple drumbeat abruptly changing tempo from 90 to 80 BPM.	17.6 s
10	abrupt tempo change	90to120	Simple drumbeat abruptly changing tempo from 90 to 120 BPM.	16.6 s
11	abrupt tempo change	60to80	Simple drumbeat abruptly changing tempo from 60 to 80 BPM.	17.4 s
12	abrupt tempo change	80to150	Simple drumbeat abruptly changing tempo from 80 to 150 BPM.	17.4 s
13	abrupt tempo change	180to120	Simple drumbeat abruptly changing tempo from 180 to 120 BPM.	13.8 s
14	abrupt tempo change	240to96	Simple drumbeat abruptly changing tempo from 240 to 96 BPM.	16.3 s
15	abrupt tempo change	85to127.5	Beat played by electronic instruments starting out in 4/4 at 85 BPM, changing to 6/4 in 127.5, with the bass and kick pattern staying the same, while the hihat and snare pattern change, enforcing the 6/4 feel.	22.6 s
16	odd measure	beat7	Drumbeat played in 7/8 measure.	15.5 s
17	odd measure	doom7	Heavy distorted guitar riff with drums in 7/8.	24.0 s
18	continuous tempo change	piano	Excerpt from Debussy's Children's Corner, L. 113: 1. Doctor Gradus ad Parnassum, played on piano with continuous tempo fluctuation.	23.3 s
19	continuous tempo change	80to200	Simple electronic drumbeat changing tempo continuously from 80 to 200 BPM.	17.2 s
20	continuous tempo change	200to80	Simple electronic drumbeat changing tempo continuously from 200 to 80 BPM.	17.2 s
21	polyrhythm	140to105	Starts with a simple electronic percussive pattern at 140 BPM, later the drum beat comes in at 105 BPM, representing the ratio of 4/3.	12.3 s
22	polyrhythm	94	A polyrhythmic drum beat playing 4 beats over 3, resetting every measure (the measures are in regular 4/4 time).	11.0 s
23	rhythmic deception	102	Normal drumbeat in 4/4 for the first half, in the second half the ride cymbal is added playing on the second 16th note of every beat, shifting the rhythmic perception.	19.6 s
24	syncopated rhythm	94	Lots of irregular syncopation, trying to cause as much confusion as possible.	21.9 s
25	odd measure	metal11	Heavy distorted guitar riff with drums in 11/8.	23.8 s

Sonification and Visualization of Annotation Data

In this section, the annotation data of my bachelor's thesis is sonified and visualized.
The data is divided into the following parts:

Average user annotations (obtained by the peaks of a Gaussian kernel density estimation function over the user annotations)
Online PLP Annotation with default parameters (Tempo Range: [30 : 240] BPM, Kernel Size: 6s)
User annotation with best F-measure
Online PLP annotation with parameters resulting in best F-measure

To listen to the sonified data, press the yellow power button in the middle of the respective trackswitcher and then press the play button or spacebar. Select the annotations you want to listen to from the tracklist below the image.

Audio 02

Starts with an arpeggio played on a synthesizer, continues in same tempo and measure with added instruments.

Audio 03

Starts with same arpeggio as the previous example, but the other instruments come in 3 8th notes later.

Audio 04

Starts with the same arpeggio, continues with different drumbeat at 2/3 of the arpeggio's tempo.

Audio 05

Simple drumbeat abruptly changing tempo from 75 to 150 BPM.

Audio 06

Simple drumbeat abruptly changing tempo from 150to 75BPM.

Audio 07

Simple drumbeat abruptly changing tempo from 75 to 112.5 BPM.

Audio 08

Simple drumbeat abruptly changing tempo from 112.5to 75 BPM.

Audio 09

Simple drumbeat abruptly changing tempo from 90 to 80 BPM.

Audio 10

Simple drumbeat abruptly changing tempo from 90 to 120 BPM.

Audio 11

Simple drumbeat abruptly changing tempo from 60 to 80 BPM.

Audio 12

Simple drumbeat abruptly changing tempo from 80 to 150 BPM.

Audio 13

Simple drumbeat abruptly changing tempo from 180 to 120 BPM.

Audio 14

Simple drumbeat abruptly changing tempo from 240 to 96 BPM.

Audio 15

Beat played by electronic instruments starting out in 4/4 at 85 BPM, changing to 6/4 in 127.5, with the bass and kick pattern staying the same, while the hihat and snare pattern change, enforcing the 6/4 feel.

Audio 16

Drumbeat played in 7/8 measure.

Audio 17

Heavy distorted guitar riff with drums in 7/8.

Audio 18

Excerpt from Debussy's Children's Corner, L. 113: 1. Doctor Gradus ad Parnassum, played on piano with continuous tempo fluctuation.

Audio 19

Simple electronic drumbeat changing tempo continuously from 80 to 200 BPM.

Audio 20

Simple electronic drumbeat changing tempo continuously from 200 to 80 BPM.

Audio 21

Starts with a simple electronic percussive pattern at 140 BPM, later the drum beat comes in at 105 BPM, representing the ratio of 4/3.

Audio 22

A polyrhythmic drum beat playing 4 beats over 3, resetting every measure (the measures are in regular 4/4 time).

Audio 23

Normal drumbeat in 4/4 for the first half, in the second half the ride cymbal is added playing on the second 16th note of every beat, shifting the rhythmic perception.

Audio 24

Lots of irregular syncopation, trying to cause as much confusion as possible.

Audio 25

Heavy distorted guitar riff with drums in 11/8.

References