| “Machine hearing” refers to the technology that the machine uses the sound sensor to collect audio data and analyze it to determine the source and change of sound.Sound event detection(SED)is one of its important contents.Voice event detection is to classify the collected voice,and judge the current event or the scene,such as judging whether the voice is "people talking","children crying" or "running water",and further judge whether it comes from the "indoor" or "outdoor" scene through the voice,which has a very broad response in the fields of smart home,security monitoring,unmanned driving,etc application prospect.With the development of neural network technology,the detection of sound events has become the choice of more and more researchers.At present,there are two difficulties in the research of this field.One is how to use weak mark or even unmarked data to learn when there is no strong mark data with the start and end time of the sound event.The other is how to improve the overlapping phenomenon of the sound event and improve the accuracy of recognition.Around the above two problems,the following research is carried out in this paper:Firstly,the neural network architecture of learning with weak label data is optimized.On the basis of extracting context information with recursive neural network(RNN),the convolutional neural network(CNN)is replaced by residual network(ResNet)to deepen the network depth and make full use of weak label data.Meanwhile,channel attention mechanism is used to enhance feature extraction and improve the recognition effect;Secondly,the mean-teacher model of semi-supervised learning framework was constructed,and three kind of data are used to train the models.Strongly labeled data,weakly labeled data and unlabeled data were used to effectively solve the problem of insufficient labeled data.At the same time,multi-channel features were used to construct multi-scale fusion features,so as to enhance features and improve the recognition effect;Finally,a visual real-time sound event detection platform is constructed and implemented on the computer.This platform is implemented in the Windows system.After the microphone device is used to collect the sound,the sound signal is fed into the signal processing and sound event recognition module,and the result is output in the display window by the python GUI module.Meanwhile,in order to enhance its real-time display,the platform provides real-time sound waveform and recording function;In order to verify the effect of model improvement and feature enhancement,experiments were conducted on DCASE 2017 task4 and DCASE 2019 task4 databases,respectively.The experimental results show that the network model and the feature enhancement scheme used in this paper have a significant effect on improving the system performance.Among them,compared with the baseline system,the sound event detection system "Res-CBAM-RNN" based on weak marker data has 0.2 lower error rate(ER)and 21% higher F-Score,which is a 0.1 percent improvement over the 2017 champion on task4.The "multi-scale feature ResNet-RNN" F-Score of sound event detection system based on semi-supervised learning increased by 11% compared with baseline and a 0.4 percent improvement over the 2018 champion on the Evaluation 2018 database. |
