A Study of Speech Emotion Change Based on the Speech Emotion Evaluation Scale (SEES) Model.

In recent years, studies on the acoustic features of emotional speech mainly focus on prosodic features, spectral based features and vowel quality features. With the deepening of the research, some of the characteristics of vocal cord vibration have been studied. Based on the speech emotion classification method in the theory of Broadcasting and Anchoring, this paper collected speech signals and phonation signals (with Electronic Glottal Graph) and extracted fundamental frequency, amplitude, open quotient, speed quotient, the first three formants, speech rate and articulation rate, and constructed a fast classification model for eight kinds of emotional speech, namely, love, hate, joy, sadness, fear, urgency, anger and normal. Through open-set test, we find that the speech emotion evaluation scale (SEES) model has high prediction accuracy and performs well in speech emotion recognition. Then, each sentence that constitutes a discourse is input into the SEES model for emotion quantification. The emotion of each sentence in the identical emotional discourse may be different, and the sentence emotion with the highest proportion is consistent with the discourse emotion in recognition. Finally, the temporal variation of emotion in discourse is observed through the SEES model, and it is found that although emotion features are distinct, the accuracy of emotion expression varies. The locations in the sentence where speech emotion recognition is high tend to be the same as the location of the focal stress in that utterance, and this may also be because the expressive purpose of the utterance and the presentation of thoughts and feelings are often realized through stress. These results may for contribute to the improvement of the speech emotion recognition rate or achieve a more natural speech emotion synthesis.