DNN Architectures and Audio Representations Comparison for Emotional Speech Classification

This work contains the results of the analysis of popular audio feature spaces for emotional speech classification accuracy. We have employed three spectro–temporal audio representations and four DNN architectures to determine the configuration most suitable to classify the speech for this task. We have tested several settings and adjustments to maximize classification accuracy. In the evaluation stage, we have used a set of emotional utterances in English and German languages. Additionally, we considered two variants when dividing the data into three groups (train, test and validation). In the first variant, the data were randomly divided by percentage. While in the second case, the sets have been additionally selected within the subgroups of individual speakers. When comparing the quality of the classification of both cases, the first option obtained better results. For both languages, we got the best performance for DenseNet121 architecture with CQT spectrogram. Interestingly, for analyzed configurations, the Mel–spectrogram was the most common representation, which received the highest classification score in 62.5% of cases.