BeatNet: CRNN and particle filtering for online joint beat downbeat and meter tracking — Proceedings of the 22nd International Society for Music Information Retrieval Conference

Claims

1. offline-comparison
   On GTZAN, BeatNet outperformed various online beat and downbeat trackers and trailed an offline joint system (Böck et al.) by 3.65% on beat tracking and 4.9% on downbeat tracking, though that offline system requires the time signature as input.

   "BeatNet underperforms [20] by 3.65% on beat tracking and by 4.9% on downbeat tracking. However, it is noted that BeatNet is an online method and it does not require the time signature input, while [20] is offline method and it requires the time signature input."

   Locator: table: 2 · Quote language: en
2. online-causal
   Offline rhythm analysis methods are typically non-causal, using future data; online methods are causal, using only past and present features, as required by real-time applications such as music accompaniment. Many online beat trackers feed a sliding window into an offline model, which causes discontinuous predictions and a priming delay; others use multi-agent models. BeatNet uses causal convolutional and recurrent layers followed by sequential Monte Carlo particle filters, and estimates meter rather than requiring a time signature.

   "Offline approaches are typically non-causal, meaning that they make predictions for a given time using data or features associated with a future time. [...] Online approaches are causal, meaning that they operate using only past and present features. These are typically desirable for human-computer interaction (HCI) systems, which must make immediate predictions, like real-time music accompaniment systems. [...] many of them [...] feed a sliding window of data into an offline model to estimate beat positions within upcoming frames. The sliding window strategy has several major drawbacks, including the discontinuity of beat predictions and the need for priming for predictions in the first window, which causes a delay [11]. Some other approaches involve inferring beat positions in real time using multi agent models [...]. we introduce an online system for joint beat, downbeat, and meter tracking, which utilizes causal convolutional and recurrent layers, followed by a pair of sequential Monte Carlo particle filters applied during inference. The proposed system does not need to be primed with a time signature in order to perform downbeat tracking, and is instead able to estimate meter and adjust the predictions over time."

   Locator: section: Abstract, 1 · Quote language: en