Densely Connected Networks with Multiple Features for Classifying Sound Signals with Reverberation.

Abstract

Listen

Translate article

In indoor environments, reverberation can distort the signalseceived by active noise cancelation devices, posing a challenge to sound classification. Therefore, we combined three speech spectral features based on different frequency scales into a densely connected network (DenseNet) to accomplish sound classification with reverberation effects. We adopted the DenseNet structure to make the model lightweight A dataset was created based on experimental and simulation methods, andhe classification goal was to distinguish between music signals, song signals, and speech signals. Using this framework, effectivexperiments were conducted. It was shown that the classification accuracy of the approach based on DenseNet and fused features reached 95.90%, betterhan the results based on other convolutional neural networks (CNNs). The size of the optimized DenseNet model is only 3.09 MB, which is only 7.76% of the size before optimization. We migrated the model to the Android platform. The modified model can discriminate sound clips faster on Android thanhe network before the modification. This shows that the approach based on DenseNet and fused features can dealith sound classification tasks in different indoor scenes, and the lightweight model can be deployed on embedded devices.