We have tried to investigate some conventional features in PSD (mean, energy, entropy) and Higher order spectra (HOS). But, the accuracy of emotional state detection is not much improved using machine learning approaches (KNN, PNN and Fuzzy). We are now thinking about the invariant features in HOS for improving the emotional state detection rate.
Hmmm (not HMM), so why would there be a non-Gaussian / non-Linear effect in emotions? So here is a speculative idea: in audio (speech, musical instruments), HOS can detect subtle changes dues to phase-coupling or non-linear effects in the sound production mechanism. If there is any sort of "jitter" effect in EEG that links emotional states to onset of such fluctuations, then HOS can provide tools that are more sensitive then normal power spectra. For instance, changes in "voicing" or "prediction gain" property of a signal can be detected using Spectral Flatness. Here is a paper about generalizing SF to non-linear effects.
http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=01316889
Dear Prof Peter H,
Our main objective is used to statistically define the changes of emotional states through EEG features. Based on the so far experimental results, we could understand that, there is some changes in EEG pattern during emotion. The HOS features are differentiating the emotions based on the information in EEG. This can be found through pattern classification results and using tomography approach. But, defining the patterns of EEG for different emotion is very challenging. Because, its purely subjective phenomenon. If you have come across with any paper work dealt with patterns based emotion detection means, pl let us know.
Dear Prof Shlomo Dubnov, Thanks for your kind answer. Let us start looking into the paper that you suggested us and hope to do well in near future..
- Badges
- Science topic