This paper explores the analysis of low-frequency components of continuous speech signals from people with Parkinson's disease, in order to detect changes in the spectrum that could be associated to the presence of tremor in the speech. Different time-frequency (TF) techniques are used for the characterization of the low frequency content of the speech signals, by paying special attention on the ability to work in non-stationary frameworks, due to the need for the analysis of long enough time segments, where the assumptions of stationary can not be met. The set of variables extracted from the TF representations includes centroids and the energy content of different frequency bands, along with entropy measures and nonlinear energy operators, which are used as features for the automatic detection of people with Parkinson's disease vs healthy controls. The discrimination capability of the estimated features is evaluated using three different classification strategies: GMM, GMM-UBM, and SVM. Furthermore, the information provided by different TF techniques is combined using a second classification stage. The results show that the changes in the low frequency components are able to discriminate between people with Parkinson's and healthy speakers with an accuracy of 77%, using one single sentence.