This paper presents a novel speech emotion recognition scheme that leverages the individuality of emotion perception. Most conventional methods simply poll multiple listeners and directly model the majority decision as the perceived emotion. However, emotion perception varies with the listener, which forces the conventional methods with their single models to create complex mixtures of emotion perception criteria. In order to mitigate this problem, we propose a majority-voted emotion recognition framework that constructs listener-dependent (LD) emotion recognition models. The LD model can estimate not only listener-wise perceived emotion, but also majority decision by averaging the outputs of the multiple LD models. Three LD models, fine-tuning, auxiliary input, and sub-layer weighting, are introduced, all of which are inspired by successful domain-adaptation frameworks in various speech processing tasks. Experiments on two emotional speech datasets demonstrate that the proposed approach outperforms the conventional emotion recognition frameworks in not only majority-voted but also listener-wise perceived emotion recognition.

To improve speech emotion recognition, a U-acoustic words emotion dictionary (AWED) features model is proposed based on an AWED. The method models emotional information from acoustic words level in different emotion classes. The top-list words in each emotion are selected to generate the AWED vector. Then, the U-AWED model is constructed by combining utterance-level acoustic features with the AWED features. Support vector machine and convolutional neural network are employed as the classifiers in our experiment. The results show that our proposed method in four tasks of emotion classification all provides significant improvement in unweighted average recall.

The majority of research in speech emotion recognition (SER) is conducted to recognize emotion categories. Recognizing dimensional emotion attributes is also important, however, and it has several advantages over categorical emotion. For this research, we investigate dimensional SER using both speech features and word embeddings. The concatenation network joins acoustic networks and text networks from bimodal features. We demonstrate that those bimodal features, both are extracted from speech, improve the performance of dimensional SER over unimodal SER either using acoustic features or word embeddings. A significant improvement on the valence dimension is contributed by the addition of word embeddings to SER system, while arousal and dominance dimensions are also improved. We proposed a multitask learning (MTL) approach for the prediction of all emotional attributes. This MTL maximizes the concordance correlation between predicted emotion degrees and true emotion labels simultaneously. The findings suggest that the use of MTL with two parameters is better than other evaluated methods in representing the interrelation of emotional attributes. In unimodal results, speech features attain higher performance on arousal and dominance, while word embeddings are better for predicting valence. The overall evaluation uses the concordance correlation coefficient score of the three emotional attributes. We also discuss some differences between categorical and dimensional emotion results from psychological and engineering perspectives.