Don't Look Back: an Online Beat Tracking Method Using RNN and Enhanced Particle Filtering

Due to inaccessibility to future data and the obligation of real-time inference, online beat tracking (OBT) has always been a challenging task. Don’t Look back! (DLB) is a novel OPT approach with an optimized structure that makes is very efficient for online performances. DLB uses a unidirectional RNN’s activations as the input to the proposed enhanced Monte-Carlo localization model to infer beat positions. It is important to note that most of the previous online methods, either apply some offline approaches to a moving window containing the past data to predict the future beat positions and/or require waiting at the beginning to get several seconds of data to initialize the parameters of the model. Meanwhile, our proposed method only uses activation of the current time frame to infer beat positions. Such a feature provides immediate beat tracking response which is a critical factor in many OPT applications. DLB outperforms state-of-the-art-models.