Method for identifying sound scenes based on CNN (convolutional neural network) and random forest classification

Abstract

The invention relates to a method for identifying sound scenes based on a CNN (convolutional neural network) and random forest classification. The method comprises the following steps of generating Mel energy spectrum and a segment sample set of the sound scene through a Mel filter; utilizing the segment sample set to perform two-phase training on the CNN, and cutting off the feature output of a whole connecting layer, so as to obtain the CNN features of the segment sample set; finally, utilizing a random forest to classify the CNN features of the segment sample set, so as to obtain a final identifying result. Proofed by related experiment results, the method has the advantage that the identification rate on an IEEE DCASE2016 sound scene evaluation data set is better than the identification rate of an MFCC-GMM (Mel frequency cepstrum coefficient-gaussian mixed model) method, and is also better than the identification rate of the existing related identifying methods.

Description

technical field [0001] The invention relates to a sound scene recognition method based on convolutional neural network and random forest classification. Background technique [0002] Sound scene recognition is to realize the perception of the sound scene by analyzing the audio signal. As one of the key links in the analysis of environmental information, it has a wide range of applications in scene recognition, foreground and background sound recognition and separation. In recent years, there have been related researches on the use of sound scene recognition to improve the terminal's autonomous perception of the scene [1][2][3] . For example, the mobile phone detects the sound of the scene and realizes automatic mute in the meeting scene; increases the volume of calls and ringtones in noisy outdoor environments; the automatic driving system analyzes the scene through the surrounding environmental sound and realizes safe driving, etc. [0003] For the recognition of sound s...

AI Technical Summary

Problems solved by technology

[0004] However, the related methods based on CNN have the following disadvantages: 1) The recognition effect depends on the length of the set segmentation, and the change of recognition rate caused by different lengths will lead to the instability of the CNN model, and the segmentation parameters need to be adjusted multiple times on the new data set, The effect produced has a certain degree of chance

2) The combination of neural networks with complex structures aggravates the complexity of the model, making it difficult to train the model

Images