Sound scene recognizing method based on label amplification and multi-spectrum fusion

Abstract

The invention discloses a sound scene recognizing method based on label amplification and multi-spectrum fusion. The method comprises the following steps: by adopting different signal processing technologies, generating various spectra for sound scene data; aiming at each spectrum, training a deep convolution neural network model which serves as a basic classification model; by utilizing the labelamplification technology, amplifying an ultra-class label for a sample, improving the original network model into a multi-task learning model by using an artificially constructed layered label, and optimizing the performances of the basic classification model; and extracting the sample features by utilizing the improved basic classification model, splicing a plurality of depth features of a voicescene file, and carrying out dimensionality reduction, thus obtaining the global features. The plurality of global features of the corresponding different spectra are fused, and an SVM classifier istrained and serves as the final classification model. According to the method, the multi-spectrum feature fusion technology is adopted, so that the recognition performance is effectively promoted; andwith the provided label amplification and model promotion method, the performances of the basic classifier can be effectively optimized, and the method can be popularized to other application researches.

Description

technical field [0001] The invention belongs to the technical field of scene recognition, and in particular relates to a sound scene recognition method based on tag amplification and multi-spectrum fusion. Background technique [0002] The sound scene recognition technology analyzes the audio data to determine the attributes, functions and uses of the space environment where the machine is located. Sound scene recognition based on convolutional neural network has become one of the most effective methods in this field. Since the sound scene dataset is labeled according to the function of the place, the problem of similarity between classes is more prominent, such as libraries and self-study classrooms, which are easy to misjudge each other. On the other hand, these data, which are inherently similar in acoustic features, are indiscriminately considered as different categories when training the network model due to different functional purposes, which hinders the network mode...

AI Technical Summary

Problems solved by technology

Similarly, Document 4 assumes that hierarchical labels already exist, and does not consider how to cluster and extract supercategory labels

Images