An Analysis of Using Fuzzy Annotations in CRNN-Based Joint Beat and Downbeat Tracking

This paper addresses joint beat and downbeat tracking by using a Convolutional Recurrent Neural Network (CRNN) trained on Mel-spectral and chroma features of musical audio signals. Since beats and downbeats occur sparsely, we extend the original beat/downbeat annotation ranges to create fuzzy annotations for a better-balanced training. We compare three fuzzy annotations generated by convolving the original annotations with triangular, Gaussian, and rectangular windows. In comparison to the baseline model trained with the original annotations, our model of using fuzzy annotations improves the beat and downbeat tracking performance on pieces with tempi slower than 80 bpm. We analyzed the activation functions from the CRNN outputs and found that the improvements are mainly due to enlarged activation functions and phase error correction. The results showed that the proposed models provide promising results on test datasets with various music styles.