Enhanced piano audio feature recognition: a novel MFCC-based method with F-HRSF and convolutional neural network

In piano audio processing and analysis, while the traditional Mel-frequency cepstral coefficients (MFCC) feature extraction method is extensively utilized in audio recognition, its recognition accuracy often falls short when applied to piano audio due to its inability to capture the intricate dynamic features of such audio fully. To address this limitation, this article enhances the MFCC feature extraction method by integrating the Fisher half rising sine function (F-HRSF) with a multilayer convolutional neural network, aiming to achieve precise recognition of piano audio features. Initially, we employ the Fisher ratio for subband screening to isolate components with strong characterization ability from the audio subband components, segmenting the MFCC features by dimensions and calculating the Fisher ratio. Subsequently, we develop an improved MFCC feature extraction method based on F-HRSF. The normalized 2D feature values obtained are then used as inputs for the multilayer hierarchical convolutional neural network. Experimental results demonstrate that the model developed in this article achieves accuracy rates of 92.15%, 92.83%, 91.57%, and 92.13% in classification accuracy, sensitivity, and specificity, respectively, on the GTZAN dataset, with a consistently stable performance in audio feature recognition. This study not only refines the MFCC feature extraction method and enhances audio feature recognition accuracy but also introduces novel ideas and approaches for piano audio processing and analysis.