Research On Power Customer Service Speech Emotion Recognition Based On Deep Learning

With the rapid growth of customer service users in the power industry,people’s demand for power supply services is increasing day by day,and the quality of customer service has become crucial to power companies.Actively responding to customer emotions is a key step in improving the quality of power customer service.However,customer service staff have a heavy workload and need to face different customers.If they cannot effectively identify and track customer emotional changes during the customer service process,the customer service cannot be eliminated.Negative emotions affect the quality of customer service.At present,there are not many studies on speech emotion recognition of power customer service.Therefore,this thesis applies the improved speech emotion recognition technology to the process of 95598 power supply customer service,which improves the quality of power supply customer service and helps power companies establish a good image.The following are some research results obtained in this thesis:(1)Aiming at the lack of emotional corpus based on the background of 95598 electric power customer service in the field of speech emotion recognition,this thesis builds a speech emotion database for power customer service.The power customer service speech script was produced by using the electric power text work order,and five performers used the script to simulate the recording in a quiet environment,and manually marked the emotion and location information of the recording.There are 4basic emotion types described,namely neutral,happy,angry and sadness.At the same time,this thesis compares and evaluates the corpus with domestic and foreign speech emotion databases.The evaluation content includes three aspects: emotion type,data scale and recording number.Afterwards,the voice data is preprocessed,including sampling quantization,pre-emphasis,frame windowing,and endpoint detection.Finally,MFCC features are extracted from the voice data to provide data support for the speech emotion recognition of electric power 95598 customer service.(2)In order to accurately identify the emotional category in the speech of electric power customer service,this thesis proposes a CNN-BLSTM-based electric power speech emotion recognition model.This model has the advantages of both CNN and BLSTM.CNN can obtain advanced features from the feature sequence,while BLSTM can learn the relationship of the context,go through the fully connected layer,and finally use the softmax function to achieve emotional classification.By setting up comparative experiments,the model is compared with a single CNN and LSTM model.The final model achieved accuracy rates of 62.19% and 63.29% on the IEMOCAP database and the self-built database,respectively.(3)On the basis of the above research,in order to further improve the recognition accuracy of the model,the CNN-BLSTM model was improved,and a power speech emotion recognition model based on CBGRU and multi-head self-attention mechanism was proposed.First,the CBGRU model is constructed by cascading.The CNN network in the model can extract the advanced features in the feature sequence.The BGRU network can learn the context-related information of the sequence,and then use the multi-head self-attention mechanism to extract the feature representation with strong discrimination,through the fully connected layer,and finally use the softmax function to achieve sentiment classification.On the one hand,by setting up four sets of comparative experiments of LSTM,GRU,CNN-BLSTM and CNN-BGRU,it is verified that the GRU network can effectively reduce the computational load of the model and shorten the training time.On the other hand,comparing the model with the CBGRU and CBGRU-attention models,the final model obtained 69.94% and 71.46%accuracy on the IEMOCAP database and the self-built database,respectively,and the recognition performance was better than that of the CBGRU and CBGRU-attention models.(4)Developed and completed a power customer service speech emotion recognition system based on CBGRU and multi-head self-attention mechanism model.The system includes a speech input module,a speech preprocessing module and a speech emotion recognition module,which can provide important technical support for speech emotion recognition of power customer service.