Determination of Low-Level Audio Descriptors of a Musical Instrument Sound Using Neural Network

Audio files and the audio channel of video files can be described with temporal, spectral, cepstral, and perceptual audio descriptors. The so-called low-level descriptors are closely related to the signal characteristics. One can discern at least three levels of extraction granularity from the signal: at any point in the signal, in small arbitrary regions (i.e., frames) and longer pre-segmented regions. Even though there are tools (e.g., MIRToolbox, Python/libROSA) available for computing these descriptors, the resulting feature vector is always redundant as it contains many high-correlated descriptors and there are some limitations connected to the performance of these tools. That is why, in this study, a method for obtaining those descriptors using Artificial Neural Network (ANN) with a deep structure (i.e., DNN) is proposed. In such a scheme, the raw audio signal representing a given musical instrument is fed to the DNN input. Such a network can be used as a standalone module or as a pre-trained part of the bigger architecture. The results of deep network performance in the context of MPEG-7 descriptor derivation are shown along with the loss function convergence and behavior.