A study on vocal style recognition using a multiscale time frequency feature fusion algorithm

In order to address the challenges facing vocal style recognition—in particular, the difficulty of characterizing transient onset, harmonic structure, and long-term melodic contours, and the lack of precision in the classification of the style, this paper proposes a novel approach to the recognition and classification of vocal styles based on multi-frequency time–frequency characteristics. In order to achieve unified modeling of acoustic information across time scales, this paper introduces 25 ms, 50 ms and 100 ms time–frequency inputs. A three-branch cooperative coding network is constructed using short-window, mid-window, and long-window inputs to represent transient onset, harmonic structure, and long-term melodic contours, respectively. Temporal attention aggregation, temperature scaling, and margin compensation are further integrated to enhance discriminative segment selection, probability calibration, and decision-boundary separation. Experiments were conducted on VocalSet and MIR-1 K using reconstructed four-class vocal style labels. The results show that the accuracy of the whole model is 91.8%, and the accuracy of cross and cross—environment testing is 88.6% and 86.9%, respectively. Compared with models that do not use short-window branches, medium-window branches, or attention aggregation mechanisms, accuracy increases by 4.5, 4.9, and 2.3 percent, respectively.