A Novel Saliency-Noise Multi-Feature Fusion for Music Signal Recognition Algorithm

The rapid development of signal processing (SP) technology, especially the continuous innovation of audio signal recognition and analysis algorithms, has led to a wide variety of application areas. Under this background, this study proposes the novel saliency-noise multi-feature fusion framework for music signal recognition algorithm, which aims to improve the recognition performance of music signals. Initially, the proposed framework extracts the emotional features of music from low-level and mid-level emotional features as the basis for overall modeling. This framework innovatively uses the eGeMAPS feature set as the low-level features, and combines the MIRToolbox tool to extract mid-and high-level music emotion features as a strong support for overall feature extraction. Subsequently, a new MFCC is designed and proposed for musical spectral modeling and generates Hilbert spectrogram through Hilbert-Huang transform. Subsequently, audio saliency detection is performed. During the detection process, the adaptive adjustment to different input signals is achieved through a two-step architecture of dynamic saliency prediction network and channel-level weighting module. Finally, the music rhythm detection is realized by constructing a histogram based on the beat position. In terms of model implementation, the sparse representation classification model is used to realize the classification and recognition of music signals. In order to test the effectiveness of the proposed algorithm, this study used a self-collected data set to experimentally verify the model. The proposed algorithm is superior to the existing Transformer and CNN models in terms of accuracy, precision, recall, and F1 value. The comparative experiment verifies the effectiveness of the proposed algorithm.