A novel deep learning-based multi-instrument recognition method for polyphonic music

Multi-instrument recognition in polyphonic music is an important research area in music information retrieval. In recent years, significant achievements have been made in this area, one of which is the attention-based method proposed by Gururani et al. However, in the attention-based method, there have been drawbacks of insufficient feature extraction and neglecting the attention weights of the frequency dimension. To solve these problems, three methods are proposed in this paper, namely Efficient Channel AttentionMIC, Spatial Group-wise Enhance AttentionMIC, and Time-Frequency AttentionMIC. The first two methods introduce a lightweight efficient channel attention mechanism and spatial group-wise enhance attention mechanism, respectively, aiming to solve the above problem of insufficient feature extraction by means of the interaction and association among different channels. The last method, namely Time-Frequency AttentionMIC, adds attention to the frequency dimension and focuses on audio features in both time and frequency domains, aiming to solve the above problem of neglecting attention weights in the frequency dimension. Experiments show that these methods not only improve the overall performance of the neural network model and the classification performance for each instrument, but also maintain the lightweight nature of the model.