Feature Extraction And Classification Based On Music Signal

Due to the difference of sound mechanism,audio processing is divided into two categories: voice and music.At present,domestic and foreign researchers mainly concentrated on modern pop music.They rarely study the primary form of Chinese ancient music(Chinese opera).Since Chinese opera has been playing an important role in Chinese traditional culture,it has important cultural significance to study Chinese opera based on computer technology.Chinese opera contains tens of thousands of arias.As a result,classifying operas manually,is not only inefficient,expensive,but also less accurate.In order to improve the management of Chinese opera,it is necessary to construct a set of automatic Chinese opera classification systems.This article proposed an algorithm to classify Chinese opera based on multi-feature fusion and extreme learning machine.Inspired by music genres classification,each aria is split into multiple segments.Nineteen features are extracted through feature engineering directly firstly.Since opera signal is a kind of temporal context sequence,the results of fast Fourier transformation(FFT)after audio frame are selected as the input,opera label of this frame is used as the target to train the long short-term memory(LSTM)neural network.Temporal context feature is extracted by the trained LSTM model.Nineteen features from feature engineering and the context feature from LSTM are combined to form the fusion feature.Finally,extreme learning machine and major voting algorithm are used to predict the opera genre of an aria.The research data include 800 arias of 8 typical genres collected from Internet.Among 8 famous Chinese traditional opera genres,temporal context feature from LSTM achieves a classification accuracy of 88.8% on average;fusion feature formed by 19 features from feature engineering and temporal context feature from LSTM finally shows an average accuracy of 92%.The experimental results demonstrated that multi-feature fusion improves classification accuracy of Chinese traditional opera genres.Temporal context feature learning by LSTM achieves better performance of temporal characteristics of the signal,which is an important supplement to features obtained from feature engineering.