Singing detection method based on multi-scale time-frequency graph parallel input convolutional neural network

Abstract

The invention discloses a singing detection method based on a multi-scale time-frequency graph parallel input convolutional neural network. Generally, in a song detection algorithm based on a convolutional neural network, a network input layer is a two-dimensional time-frequency graph matrix. The method comprises the following steps: firstly, according to the multi-scale characteristic of a music signal, generating a plurality of two-dimensional time-frequency graph matrixes with different scales by adjusting the window length of short-time Fourier transform; and then sending the plurality of time-frequency diagrams to a convolutional neural network in a parallel multi-channel mode, so that a neuron receptive field of the convolutional neural network can simultaneously observe multi-scale information of a music signal, thereby enhancing the time-frequency diagram feature extraction and resolution capability of neurons, and improving the overall performance of singing detection.

Description

technical field [0001] The invention relates to the technical field of music artificial intelligence, in particular to a singing voice detection method based on a multi-scale time-frequency graph input in parallel to a convolutional neural network. Background technique [0002] Regarding the background technology of singing voice inspection, the applicant has developed a singing voice detection method based on squeeze and incentive residual network (application number: CN202010164594.5) and a singing voice detection method based on dot product self-attention convolutional neural network (patent number : ZL202110192300.4) are described. Singing Voice Detection (SVD) is a process of judging whether each small piece of audio in digital music contains singing voices, and its detection accuracy is generally between 50-200 milliseconds. Singing detection is an important basic work in the field of Music Information Retrieval (MIR). Many other research directions, such as singer re...

AI Technical Summary

Problems solved by technology

The invention implicitly extracts singing features at different levels through the deep residual network, and can use the adaptive attention characteristics of the embedded extrusion and excitation modules to judge the importance of these features. Under the JMD data set, the depths are respectively In the case of 14, 18, 34, 50, 101, 512, and 200, the average detection accuracy rate is 88.19, and the effect still needs to be improved

In addition, the method of stacking networks in this invention consumes a lot of computing resources and increases training time.

Images