Abstract: Speech emotion recognition(SER) plays an important role in human-computer interaction systems.In order to solve the problems that the decision making process of existing deep learning models is opaque in the SER task,and the traditional concept-cognitive learning(CCL) is susceptible to noise interference and concept drift when processing incremental data,a three-way weighted fuzzy concept-cognitive classification framework that leverages extremely randomized trees(3WERT-WFCCL) is proposed.In the feature processing,Whisper is used to extract high dimensional speech features,and a multi-layer perceptron(MLP) is used for hierarchical abstract representation.In the cognitive learning stage,the extremely randomized trees(ERT) algorithm is introduced to calculate the importance of features to realize the automated quantitative allocation of attribute weights,and the three-way decision fault tolerance threshold parameter is embedded in the cognitive operator to construct a positive and negative two-way cognitive mechanism.In the face of incremental data,the model divides the new samples into a positive region,a boundary region and a negative region according to the feature identification distance,and adopts a robust strategy that only uses the positive region samples to update the concept,which effectively resists the noise interference.On the SAVEE dataset with more complex feature boundaries,the robust update strategy improves the accuracy by 0.16 percentage points compared with the global update strategy.Experiments on two public datasets EmoDB and SAVEE show that 3WERT-WFCCL is superior to the existing baseline methods in multiple key evaluation indicators.Compared with the baseline models Logistic Regression(LR) with the best performance on each data set,the accuracy of the proposed algorithm is increased by 1.53 percentage points and 0.62 percentage points respectively,and the F1 score is increased by 1.28 percentage points and 0.40 percentage points respectively.Experimental results verify the effectiveness of the three-way decision mechanism,which provides a new method for constructing SER models with high classification accuracy,strong noise robustness and logical interpretability.

Key words: Concept-cognitive learning, Three-way decision, Speech emotion recognition, Granular computing, Formal concept analysis