Studies have shown that the performance of speech recognition algorithms severely degrade due to the presence of task and emotional induced stress in adverse conditions. This paper addresses the problem of detecting the presence of stress in speech by analyzing nonlinear feature characteristics in specific frequency bands. The framework of the previously derived Teager Energy Operator(TEO) based feature TEO-CB-AutoEnv is used. A new detection scheme is proposed based on weighted TEO features derived from critical bands frequencies. The new detection framework is evaluated on a military speech corpus collected in a Soldier of the Quarter (SOQ) paradigm. Heart rate and blood pressure measurements confirm subjects were under stress. Using the traditional TEO-CB-AutoEnv feature with an HMM trained stressed speech classifier, we show error rates of 22.5% and 13% for stress and neutral speech detection. With the new weighted sub-band detection scheme, detection error rates are reduced to 4.7% and 4.6% for stress and neutral detection, a relative error reduction of 79.1% and 64.6% respectively. Finally we discuss issues related to generation of stress anchor models and speaker dependency.