Semi-supervised music main melody extraction method

Abstract

The invention discloses a semi-supervised music main melody extraction method, which comprises the following steps of: performing normalization, resampling and filtering preprocessing on an input audio signal to obtain an audio signal for simulating auditory characteristics of a human ear; carrying out constant Q spectrum transformation on the audio signal to obtain a variable resolution frequencyspectrum signal with logarithmically distributed frequency, aggregating adjacent multiple frames of amplitude spectrums to obtain a feature vector, constructing an input vector set of an extreme learning machine according to the feature vector, and obtaining an output vector set according to the training set; carrying out parameter training on the extreme learning machine, and carrying out melodypitch coarse estimation by utilizing an extreme learning machine network; searching for a spectrum peak within a rough estimated 2 / 3 half-tone range of the melody pitch of each frame, taking the frequency corresponding to the spectrum peak as the melody pitch of the frame to be output, and finely adjusting the melody pitch.

Description

technical field [0001] The invention belongs to the field of audio signal processing, in particular to a semi-supervised music main melody extraction method. Background technique [0002] The features used in the existing main melody extraction methods are divided into two categories: the first category is low-level artificial features, which are set by researchers based on signal characteristics, and their performance is significantly reduced when dealing with complex music with multiple sound sources. The second type of features is advanced self-learning features, which are obtained by the algorithm on the basis of the data set, and its performance depends heavily on the capacity and diversity of the training set. Recently, deep learning has provided a new solution for music main melody extraction, but it takes a long time to train to obtain appropriate network parameters, and the capacity and diversity of the training set are still important factors restricting its perfor...

AI Technical Summary

Problems solved by technology

[0002]The features used in the existing main melody extraction methods are divided into two categories: the first category is low-level artificial features, which are set by researchers according to the characteristics of the signal. The performance of complex music with multiple sound sources drops significantly, and its generalization is poor; the second type of features is advanced self-learning features, which are obtained by the algorithm on the basis of data sets, and its performance depends heavily on the capacity of the training set and diversity

Recently, deep learning has provided a new solution for the extraction of the main theme of music, but it takes a long time to train to obtain appropriate network parameters, and the capacity and diversity of the training set are still important factors restricting its performance.

However, existing annotated datasets for theme extraction are still lacking, which limits the performance of such methods.

Images