Music Melody Extraction Algorithm Based on Multi-Layer Perceptron Neural Network
Mengyuan Sun · 2024
Current melody extraction methods face challenges from the complexity of audio signal processing and multi-track interference. This study proposes a music melody extraction algorithm based on a multilayer perceptron neural network. The algorithm uses a multilayer perceptron to build a model. The input layer processes the audio signal features through short-time Fourier transform (STFT), taking into account spectral and time domain information. The model is designed with three hidden layers, each containing a large number of neurons, and nonlinear transformation is performed through the ReLU activation function. At the same time, the dropout technology is used to control overfitting. The output layer uses the soft max activation function to generate the probability distribution of the melody signal. The highest accuracy of the melody extraction task reached 95%, and the precision and recall rates were 87%. The melody extraction algorithm based on the multilayer perceptron performs well in processing complex audio signal tasks, providing an effective solution for the field of melody extraction.