Complex sound recognition method based on one-dimensional convolutional neural network

Abstract

The invention discloses a complex sound recognition method based on a one-dimensional convolutional neural network. The method comprises the steps: a complex sound is processed through a random completion algorithm, and original data is filled into the same length for input of the one-dimensional convolutional neural network; and a pre-emphasis module and a simplified attention mechanism module are embedded in a basic framework of the one-dimensional convolutional neural network, wherein the pre-emphasis module is arranged at an input part of the one-dimensional convolutional neural network and is used for pre-emphasis on input data and participating in network model optimization; and the simplified attention mechanism module is arranged in a deep layer of the one-dimensional convolutional neural network, and global features with attention are obtained by using global average pooling and a sigmoid function. Through the method, the network model is optimized, and a good identification effect is obtained.

Description

technical field [0001] The invention belongs to the technical field of audio processing, and relates to complex sound recognition technology, in particular to a complex sound recognition method based on a one-dimensional convolutional neural network. Background technique [0002] Complex sounds refer to non-linguistic sounds in the environment. The sound sources are complex and diverse, the signal itself is non-stationary and often accompanied by extremely disturbing background noise, etc., making the sound features of different sound scenes not obvious enough or the feature similarity is very high. , complex sound recognition can automatically identify specific types of complex sounds in the environment, such as children playing, car horns and street music. In the field of sound classification, such as speech classification and music classification have achieved very high accuracy, but in the field of complex sound recognition due to the non-stationary signal itself, the sp...

AI Technical Summary

Problems solved by technology

The first is to directly use the original signal for network training. The advantage of this method is that it does not require manual feature extraction of the signal, which greatly simplifies the operation process, and the model is simple and convenient to popularize; the second method is to process the original data and manually Extract some features of the sound signal, such as spectrogram and Mel frequency cepstral coefficient, etc. This method has the advantage of high accuracy for some data sets, but the generalization ability is poor, and it is difficult for the model to change the subsequent recognition results

The third is a multi-input complex network, which uses the original sound signal and artificially extracted features as the input part of the network. Defects, but this type of model is more complicated, has high requirements on the hardware of the platform, and is inconvenient to apply

However, since the deep learning model is difficult to effectively extract the features of the original signal, and the model proposed by the existing technology is relatively complex, further optimization is required.

Therefore, solving complex sound problems based on the original audio signal is a very big challenge

There will be many audio durations in the data set that are very different from the target length. For example, the actual duration is 1 second and the target length is 4 seconds. Obviously, the cubic spline interpolation is no longer suitable for this situation, and the zero padding method is too simple, and a lot of data will be lost. information, and the more zeros are filled, the more effective information may be concealed

Images