Robust Speech Emotion Recognition Using Log Frequency Power Ratio

Nowadays, recognizing human emotion is an important issue in the human-robot interaction field. Accordingly, in this paper a novel feature for emotion recognition is proposed. In particular, the emotion recognition system introduced in this study uses information extracted from human speech only. In addition, during speech emotion recognition, the intensity of the speech, the volume of the microphone and distance between the recognition system and the speaker can cause several problems and influence the performance of system. Therefore, a feature which is robust in amplifying the speech signal is included. This feature, the log frequency power ratio, is related to the ratio of the frequency power of each frequency band in the filter bank, which is modeled as a human audible system. It shows a better performance in terms of amplifying compared to a feature only related to absolute power. To evaluate the new feature performance, it is compared to the performance of an existing feature, LFPC. In conclusion, the new feature does not show an improvement in a clean environment, but it does perform better in terms of signal magnification. Therefore, it is expected that this feature could be more useful for an emotion recognition system in actual environments