DNN-LSTM-CRF Model for Automatic Audio Chord Recognition
Shota Nakayama, Shuichi Arai · 2018
Automatic Chord recognition systems use timewise models to post-process frame-wise chord predictions from acoustic models. In this paper, we propose a DNN - LSTM - CRF model for chord recognition. Deep learning has become widespread in chord recognition and has good effects for improving recognition accuracy. It is common to perform chord classification with CNN and perform final chord recognition using RNN, CRF or the like. However, CRF commonly used for post filtering considers chord information at the previous time only. Since there is a progress pattern for chord progression in music We thought that estimation accuracy could be improved by using chord information at a previous time. Therefore, we focused on LSTM (Long Short - Term Memory). We think that improvement of estimation accuracy can be expected by considering chord information of many times in LSTM and further using chord information of previous time in CRF. Although a hybrid model of RNN and CRF has been proposed, RNN has a kind and it is necessary to select an appropriate RNN. Therefore, we conducted comparison study using three types of RNN, LSTM, GRU and Bi-LSTM. Our results showed the effectiveness of RNN in chord recognition using deep learning. This discovery constitutes a further step towards the development of a chord recognition system.