Fractal-Based Speech Analysis for Emotional Content Estimation

Speech emotional content estimation is still a challenge for building robust human-machine interaction systems. Accuracy of emotion estimation depends upon the corpus used for training and the acoustic features employed for modelling the speech signal. Generally, emotion estimation is computationally expensive, and hence, there is a need of developing alternative techniques. In this paper, a low complexity fractal-based technique has been explored. Our hypothesis is that fractal analysis would provide better emotional content estimation because of the nonlinear nature of the speech signals. Fractal analysis involves two important parameters, i.e. fractal dimension and loop area. Fractal dimension has been computed using the Katz algorithm. The investigations using a GMM-based model show that the proposed technique is capable of identifying the emotional content within the given speech signals reliably and accurately. Further, the technique is robust in the sense that it can bear the noise level in the signal up to 10 dB. The analysis also shows that the technique is gender insensitive. The scope of the investigations presented here is limited to phonemic-level analysis, although the technique works efficiently with speech phrases as well.