TS-Resformer: a model based on multimodal fusion for the classification of music signals.

The number of music of different genres is increasing year by year, and manual classification is costly and requires professionals in the field of music to manually design features, some of which lack the generality of music genre classification. Deep learning has had a large number of scientific research results in the field of music classification, but the existing deep learning methods still have the problems of insufficient extraction of music feature information, low accuracy rate of music genres, loss of time series information, and slow training. To address the problem that different music durations affect the accuracy of music genre classification, we form a Log Mel spectrum with music audio data of different cut durations. After discarding incomplete audio, we design data enhancement with different slicing durations and verify its effect on accuracy and training time through comparison experiments. Based on this, the audio signal is divided into frames, windowed and short-time Fourier transformed, and then the Log Mel spectrum is obtained by using the Mel filter and logarithmic compression. Aiming at the problems of loss of time information, insufficient feature extraction, and low classification accuracy in music genre classification, firstly, we propose a Res-Transformer model that fuses the residual network with the Transformer coding layer. The model consists of two branches, the left branch is an improved residual network, which enhances the spectral feature extraction ability and network expression ability and realizes the dimensionality reduction; the right branch uses four Transformer coding layers to extract the time-series information of the Log Mel spectrum. The output vectors of the two branches are spliced and input into the classifier to realize music genre classification. Then, to further improve the classification accuracy of the model, we propose the TS-Resformer model based on the Res-Transformer model, combined with different attention mechanisms, and design the time-frequency attention mechanism, which employs different scales of filters to fully extract the low-level music features from the two dimensions of time and frequency as the input to the time-frequency attention mechanism, respectively. Finally, experiments show that the accuracy of this method is 90.23% on the FMA-small dataset, which is an improvement in classification accuracy compared with the classical model.