In this paper, we discuss a new Principal Component Analysis (PCA)-based manifold representation for visual speech recognition. In this regard, the real time input video data is compressed using Principal Component Analysis and the low-dimensional points calculated for each frame define the manifold. Since the number of frames that form the video sequence is dependent on the word complexity, in order to use these manifolds for visual speech classification it is required to re-sample them into a fixed pre-defined number of key-points. These key-points are used as input for a Hidden Markov Model (HMM) classification scheme. We have applied the developed visual speech recognition system to a database containing a group of English words and the experimental data indicates that the proposed approach is able to produce accurate classification results.