Music instrument recognition using deep convolutional neural networks

Musical instruments identification in polyphonic is a challenge in music information retrieval. In proposed work, a deep convolution neural network framework for predominant instrument recognition in real-world polyphonic music is accomplished. The network is trained on fixed-length music with a labeled predominant instrument and estimate an arbitrary number of instruments from an audio signal with variable length. The Mel spectrogram representation is used to map audio data into the matrix format. This work used eight layer convolution neural network for instrument recognition. ReLu activation function is used for the scaling of training data and introduces non-linearity in the network. At each layer, Max Pooling function is used for the dimension reduction. For the regularization, dropout is used which prevent the output from getting overfitting. The Softmax function gives the probability of particular instruments. The research excellent result with 92.8% accuracy. © 2019, Bharati Vidyapeeth's Institute of Computer Applications and Management.