Melody extraction aims to produce a sequence of frequency values corresponding to the pitch of the dominant melody from a musical recording, comprising a large variety of algorithms spanning a wide range of techniques. In this paper, a novel DNN-LSTM based architecture is proposed for melody extraction. Melody extraction is regard as a composition of pitch estimation and voicing detection. This paper present a multi-task learning approach so as to perform the two tasks simultaneously, which proves to help the model obtain higher accuracy and better generalization ability. Experiments on public datasets show that the proposed model is capable of modeling temporal dependencies, and have a comparable result to the state-of-the-art methods.
展开▼
