Scalogram vs Spectrogram as Speech Representation Inputs for Speech Emotion Recognition Using CNN

Speech Emotion Recognition (SER) focuses on understanding the human emotion in a given speech utterance using its acoustic and/or linguistic features. This paper presents a comparison between two speech representation inputs for SER: spectrograms and scalograms. Speech signals from four databases (Emo-DB, RAVDESS, SAVEE, and a mix of all three) were converted into each type of representation and were used to train variations of a convolutional neural network (CNN) VGG16 Model-3. Results show that the scalogram-based models have a higher mean f1-score compared to the spectrogram-based models; however, further analysis indicate that the difference is statistically insignificant at a 95% confidence level. In conclusion, spectrograms and scalograms have statistically the same performance on the systems presented.