Research On Acoustic Scene Classfication Method Based On Subspectrogram

Acoustic Scene Classification(ASC)is one of the main research contents of computational auditory scene analysis.Its aim is to classify a given audio into one of the predefined acoustic scene categories.Acoustic scene classification has received more and more attention due to its large application value.The development of the Internet of Things has provided ASC with application scenarios,and the development of artificial intelligence has provided strong technical support for its research.With the widespread application of neural networks,Convolutional Neural Network(CNN)has gradually become the mainstream method in this field.Correspondingly,the Mel spectrogram has become the most commonly used feature.Due to the different distribution of sound frequencies in acoustic scenes,some researchers have begun to use sub-spectrogram-based methods to classify acoustic scenes.However,this type of method cannot obtain sufficiently good global features during training,which affects the performance of the model.To this end,this paper conducts further research on the sub-spectrogram-based acoustic scene classification method.The main research contents and innovations are as follows:1.Two-stage acoustic scene classification method based on spectrogram subbandsThis paper chooses the Sub Spectral Net model based on sub-spectrogram as the baseline system.To overcome its limitations in training methods,a two-stage acoustic scene classification method based on spectrogram subbands is proposed.The first stage is the feature optimization stage,which optimizes the features of each spectrogram subband,hoping to obtain better local features;the second stage is the global classification stage,the local features obtained in the first stage are spliced to construct global features,and then complete the classification of the input samples.2.Two-stage acoustic scene classification method based on spectrogram subregionsIn an acoustic scene,many key sounds may only be concentrated in certain small time periods.If the entire frequency band is processed,too much noise will be introduced,which will cause interference to the classification results.To solve this problem,an acoustic scene classification method based on spectrogram sub-regions is proposed.First,the features extracted from the spectrogram sub-bands are further divided in the time dimension to divide the spectrogram sub-regions,and then each spectrogram sub-region is used for classification,and then according to the contribution of each spectrogram sub-region to the scene,different weights are used to weight the classification results of each spectrogram sub-region,so as to achieve better classification accuracy.3.Acoustic scene classification based on the combination of sub-frequency spectrogram and complete spectrogramTo overcome the problem that the above methods have poor classification results in certain categories,an acoustic scene classification method based on the combination of sub-spectrogram and complete spectrogram is proposed.While using the sub-spectrogram,the complete spectrogram is introduced to classify the acoustic scene,and finally the output of the sub-network is weighted to obtain the highest classification accuracy rate of 72.3%.