There is no definite answer to that question since different workflows are applied to different tasks. For example in some applications you might want to normalize your input data samples or perform some kind of Automatic Gain Control (i.e. you work in fixed point and even for some low signal values you want your FFT to have good resolution). On the contrary you might be using some energy features, such as very simple energy based Activity Detection, thus normalisation of your input samples will mess your results. Therefore you must use your common sense and think (or test) whether your features are dependent on input amplitude.
Other thing that you probably had in mind, is to normalize the features itself. This is called a whitening and is used for example in speech recognition for removal of channel effects. It is done by removing mean of your features and (if you wish) normalizing by the variance. This make your features to be in the same space.
To wrap up, it's up to you. Train/test both systems with and without any type of normalization and observe the performance.