Features extracted with optimized wavelets were compared with standard methods for a Brain-Computer Interface driven by non-motor imagery tasks. Two non-motor imagery tasks were used, Auditory Imagery of a familiar tune and Spatial Navigation Imagery through a familiar environment. The aims of this study were to evaluate which method extracts features that could be best differentiated and determine which channels are best suited for classification. EEG activity from 18 electrodes over the temporal and parietal lobes of nineteen healthy subjects was recorded. The features used were autoregressive and reflection coefficients extracted using autoregressive modeling with several model orders and marginals of the wavelet spaces generated by the Discrete Wavelet Transform (DWT). An optimization algorithm with 4 and 6 taps filters and mother wavelets from the Daubechies family were used. The classification was performed for each single channel and for all possible combination of two channels using a Bayesian Classifier. The best classification results were found using the marginals of the Optimized DWT spaces for filters with 6 taps in a 2 channels classification basis. Classification using 2 channels was found to be significantly better than using 1 channel (p<<0.01). The marginals of the optimized DWT using 6 taps filters showed to be significantly better than the marginals of the Daubechies family and autoregressive coefficients. The influence of the combination of number of channels and feature extraction method over the classification results was not significant (p=0.97).