Beat this! Accurate beat tracking without DBN postprocessing — Proceedings of the 25th International Society for Music Information Retrieval Conference

Claims

1. dataset-quality
   The authors found multiple annotation problems in beat-tracking datasets, including GTZAN: some pieces have clear annotation errors, some would be hard even for experts to annotate uniquely, and some question whether there is a beat to track at all. They called for correcting existing datasets and producing annotated data for underrepresented genres.

   "While exploring the datasets, we found multiple problems in the annotations, and we think this hinders the development of better models, especially for downbeat predictions. Even the GTZAN dataset, which is commonly used for evaluation, is not immune to quality problems. Some of them are evident and not debatable [...]. Furthermore, there are pieces where even for experts it would be hard to agree on a unique beat and downbeat annotation [...]. Finally, some pieces question the primary assumption of beat tracking, i.e., that there is a beat/downbeat to track [...]. correcting the existing commonly used datasets, and producing new annotated data for underrepresented genres is a crucial step for further development."

   Locator: section: 4.5, 5 · Quote language: en
2. datasets
   Beat This! was trained and validated on Simac, SMC, Hainsworth, Ballroom, HJDB, Beatles, Harmonix, RWC (classical, pop, royalty-free, jazz), TapCorrect, JAAH, Filosax, ASAP, Groove MIDI, GuitarSet, and Candombe (4,556 tracks), with GTZAN used for testing only.

   "We train and validate with several datasets: Simac [26], SMC [27], Hainsworth [28], Ballroom [26, 29], HJDB [30], Beatles [31], Harmonix [32], RWC [33, 34] (classical, pop, royalty-free, and jazz), TapCorrect [35], JAAH [36], Filosax [37], ASAP [38], Groove MIDI [39], GuitarSet [40], Candombe [41]. [...] resulting in a total of 4556 tracks. [...] We use the GTZAN [13] dataset [...] for testing only."

   Locator: section: 4.1 · Quote language: en
3. dbn-limitations
   The dynamic Bayesian network (DBN) postprocessing proposed by Böck et al. is the most widely used postprocessing for neural beat trackers and was employed by all recent high-accuracy models; it constrains tempo to a fixed range and beats per bar to a short list, so it is bound to fail on pieces with time-signature changes, tempos outside the range, or unsupported numbers of beats per bar.

   "The most widely used postprocessing technique is the Dynamic Bayesian Network (DBN) in the form proposed by Böck et al. [9]. [...] We argue that the DBN is a problematic component because it is inherently bound to fail for several music pieces: pieces with time signature changes, pieces whose tempo falls outside of the tempo range (or that slow down/speed up outside the tempo range), and pieces whose number of beats per measure are not included in the list of supported values. [...] Remarkably, we do not use the DBN postprocessing, which was employed by all recent high-accuracy models."

   Locator: section: 1, 5 · Quote language: en
4. system-and-results
   Beat This! seeks generality by training on multiple diverse datasets and removing DBN postprocessing, using an architecture alternating convolutions with transformers; it surpassed the state of the art in F1 score but performs worse on continuity metrics and can still fail on difficult and underrepresented genres. The authors released the model, code, and preprocessed datasets.

   "We achieve generality by training on multiple datasets – including solo instrument recordings, pieces with time signature changes, and classical music with high tempo variations – and by removing the commonly used Dynamic Bayesian Network (DBN) postprocessing, which introduces constraints on the meter and tempo. For high accuracy, among other improvements, we develop a loss function tolerant to small time shifts of annotations, and an architecture alternating convolutions with transformers either over frequency or time. Our system surpasses the current state of the art in F1 score despite using no DBN. However, it can still fail, especially for difficult and underrepresented genres, and performs worse on continuity metrics, so we publish our model, code, and preprocessed datasets, and invite others to beat this."

   Locator: section: Abstract · Quote language: en