Skip That Beat: Augmenting Meter Tracking Models for Underrepresented Time Signatures

Giovana Morais
MARL

Brian McFee
MARL-CDS

Magdalena Fuentes
MARL-IDM

Paper Code

Abstract

Beat and downbeat tracking models are predominantly developed using datasets with music in 4/4 meter, which decreases their generalization to repertories in other time signatures, such as Brazilian samba which is in 2/4. In this work, we propose a simple augmentation technique to increase the representation of time signatures beyond 4/4, namely 2/4 and 3/4. Our augmentation procedure works by removing beat intervals from 4/4 annotated tracks. We show that the augmented data helps to improve downbeat tracking for underrepresented meters while preserving the overall performance of beat tracking in two different models. We also show that this technique helps improve downbeat tracking in an unseen samba dataset.

Dataset biases

Augmentation Examples

Augmentation Procedure

Augmentation Procedure

Beatles

Original (4/4)

Augmented (2/4)

Augmented (3/4)


GTZAN

Original (4/4)

Augmented (2/4)

Augmented (3/4)


RWC Jazz

Original (4/4)

Augmented (2/4)

Augmented (3/4)


RWC Classical

Original (4/4)

Augmented 2/4

Augmented (3/4)


Training Data Distribution

Training Data Distribution

Complete Results

BayesBeat

Beat evaluation (Test set)

  training data F CMLc CMLt AMLc AMLt
2/4 B 0.582 0.241 0.322 0.509 0.679
  AugF 0.573 0.223 0.298 0.529 0.696
  AugS 0.580 0.259 0.323 0.526 0.698
3/4 B 0.608 0.344 0.444 0.455 0.644
  AugF 0.642 0.390 0.588 0.502 0.688
  AugS 0.631 0.388 0.489 0.496 0.679
4/4 B 0.842 0.673 0.714 0.826 0.882
  AugF 0.842 0.673 0.714 0.826 0.883
  AugS 0.817 0.644 0.683 0.810 0.872

Downbeat evaluation (Test Set)

  training data F CMLc CMLt AMlc AMLt
2/4 B 0.376 0.182 0.218 0.407 0.484
  AugF 0.401 0.202 0.242 0.408 0.488
  AugS 0.418 0.182 0.213 0.422 0.517
3/4 B 0.193 0.032 0.047 0.068 0.130
  AugF 0.424 0.313 0.345 0.401 0.480
  AugS 0.293 0.076 0.087 0.151 0.188
4/4 B 0.609 0.559 0.572 0.782 0.798
  AugF 0.602 0.464 0.477 0.745 0.764
  AugS 0.512 0.282 0.292 0.613 0.631

BRID: Beat and Downbeat evaluation

  training data F CMLc CMLt AMLc AMLt
Beat B 0.612 0.147 0.151 0.690 0.709
  AugF 0.606 0.147 0.150 0.689 0.703
  AugS 0.575 0.124 0.127 0.627 0.644
Downbeat B 0.068 0.069 0.072 0.272 0.275
  AugF 0.137 0.007 0.007 0.253 0.275
  AugS 0.224 0.030 0.038 0.126 0.160


TCN-PP

Beat evaluation (Test set)

  training data F CMLc CMLt AMLc AMLt
2/4 B 0.553 0.030 0.090 0.231 0.417
  AugF 0.547 0.029 0.105 0.189 0.382
  AugS 0.526 0.022 0.084 0.164 0.366
3/4 B 0.680 0.167 0.392 0.220 0.498
  AugF 0.706 0.244 0.451 0.288 0.530
  AugS 0.672 0.181 0.385 0.236 0.501
4/4 B 0.848 0.495 0.618 0.602 0.752
  AugF 0.877 0.567 0.694 0.635 0.777
  AugS 0.831 0.464 0.586 0.573 0.722

Downbeat evaluation (Test Set)

  training data F CMLc CMLt AMLc AMLt
2/4 B 0.352 0.003 0.012 0.040 0.097
  AugF 0.357 0.002 0.012 0.040 0.125
  AugS 0.328 0.002 0.010 0.035 0.091
3/4 B 0.349 0.000 0.002 0.194 0.250
  AugF 0.369 0.001 0.003 0.275 0.335
  AugS 0.341 0.000 0.001 0.208 0.254
4/4 B 0.369 0.000 0.000 0.103 0.122
  AugF 0.361 0.000 0.000 0.051 0.070
  AugS 0.324 0.000 0.000 0.014 0.019


TCN-DBN

Beat evaluation (Test set)

  training data F CMLc CMLt AMLc AMLt
2/4 B 0.643 0.321 0.425 0.522 0.715
  AugF 0.641 0.373 0.482 0.525 0.704
  AugS 0.571 0.251 0.337 0.439 0.611
3/4 B 0.716 0.472 0.606 0.548 0.720
  AugF 0.720 0.475 0.629 0.539 0.737
  AugS 0.723 0.499 0.665 0.539 0.728
4/4 B 0.900 0.786 0.827 0.887 0.933
  AugF 0.906 0.810 0.855 0.891 0.945
  AugS 0.877 0.740 0.777 0.858 0.914

Downbeat evaluation (Test Set)

  training data F CMLc CMLt AMLc AMLt
2/4 B 0.447 0.041 0.067 0.377 0.482
  AugF 0.439 0.024 0.046 0.400 0.507
  AugS 0.389 0.023 0.044 0.297 0.382
3/4 B 0.422 0.035 0.051 0.616 0.669
  AugF 0.410 0.014 0.025 0.617 0.676
  AugS 0.403 0.015 0.024 0.615 0.674
4/4 B 0.353 0.000 0.000 0.023 0.025
  AugF 0.371 0.000 0.000 0.045 0.050
  AugS 0.347 0.000 0.000 0.058 0.061


Cite