Gender and affect recognition based on GMM and GMM-UBM modeling with relevance MAP estimation

The paper presents our efforts in the Gender Sub-Challenge and the Affect Sub-Challenge of the INTERSPEECH 2010 Paralinguistic Challenge. The system for the Gender Sub-Challenge is based on modeling the Mel-Frequency Cepstrum Coefficients using Gaussian mixture models, building a separate model for each of the gender categories. For the Affect Sub-Challenge we propose a modeling schema where a universal background model is first trained an all the training data and then, employing the maximum a posteriori estimation criteria, a new feature vector of means is produced for each particular sample. The feature set used is comprised of low level descriptors from the baseline system, which in our case are split into four subsets, and modeled by its own model. Predictions from all subsystems are fused using the sum rule fusion. Aside from the baseline regression procedure, we also evaluated the Support Vector Regression and compared the performance. Both systems achieve higher recognition results on the development set compared to baseline, but in the Affect Sub-Challenge our system’s cross correlation is lower than that of the baseline system, although the mean linear error is slightly superior. In the Gender Sub-Challenge the unweighted average recall on the test set is 82.84%, and for the Affect Sub-Challenge the cross-correlation on the test set is 0.39 with mean linear error of 0.143.