Joint Beat and Downbeat Tracking with Recurrent Neural Networks — Proceedings of the 17th International Society for Music Information Retrieval Conference (ISMIR 2016)

Claims

1. downbeats-applications
   Beats and downbeats form the lower levels of music's metrical hierarchy; downbeats have received less research than beats but matter for structural segmentation, music analysis, and automated DJ mixing.

   "The lower levels of this hierarchy are defined by the beats and downbeats which define the metrical structure of a musical piece. While considerable amount of research focused on finding the beats in music, far less effort has been made to track the downbeats, although this information is crucial for a lot of higher level tasks such as structural segmentation and music analysis and applications like automated DJ mixing."

   Locator: section: 1. Introduction · Quote language: en
2. evaluation-practice
   Downbeat tracking is usually evaluated with F-measure at a ±70 ms tolerance; comparisons between systems are often impossible because of differing metrics and datasets; test sets include the SMC set of hard-to-track pieces and Carnatic, Cretan, and Turkish music.

   "In line with almost all other publications on the topic of downbeat tracking, we report the F-measure (F1) with a tolerance window of ±70 ms. [...] We do not report scores for all sets used for training, since comparisons with other works are often not possible due to different evaluation metrics and/or datasets. [...] the SMC set (built specifically to comprise hard-to-track musical pieces) [...] For non-western music we use the Carnatic, Cretan, and Turkish datasets"

   Locator: section: 3. Evaluation · Quote language: en
3. rnn-dbn
   A recurrent neural network on magnitude spectrograms models multiple metrical levels to distinguish beats from downbeats, and a dynamic Bayesian network aligns them to bars of variable length, achieving state-of-the-art results across many genres.

   "A recurrent neural network operating directly on magnitude spectrograms is used to model the metrical structure of the audio signals at multiple levels and provides an output feature that clearly distinguishes between beats and downbeats. A dynamic Bayesian network is then used to model bars of variable length and align the predicted beat and downbeat positions to the global best solution. We find that the proposed model achieves state-of-the-art performance on a wide range of different musical genres and styles."

   Locator: section: Abstract · Quote language: en