It is now widely acknowledged, among communities of researchers and engineers of very different horizons, that speckle interferometry (SI) offers powerful techniques to characterize mechanical rough surfaces with a submicronic accuracy in static or quasi-static regime, when small displacements are involved (typically several microns or tens of microns). The issue of dynamic regimes with possibly large deformations (typically several hundreds of microns) is still topical and prevents an even more widespread use of speckle techniques. This is essentially due to the lack of efficient processing schemes able to cope with non-stationary AM-FM interferometric signals. In addition, decorrelation-induced phase errors represent an hindrance to accurate measurement when such large displacements and classical fringe analysis techniques are considered. This work is an attempt to address those issues and to endeavor to make the most of speckle interferometry signals. Our answers to those problems are located on two different levels. First of all, we adopt the temporal analysis approach, i.e. the analysis of the temporal signal of each pixel of the sensor area used to record the interferograms. A return to basics of phase extraction is operated to properly identify the conditions under which the computed phase is meaningful and thus give some insight on the physical phenomenon under analysis. Due to their intrinsic non-stationary nature, a preprocessing tool is missing to put the SI temporal signals in a shape which ensures an accurate phase computation, whichever technique is chosen. This is where the Empirical Mode Decomposition (EMD) intervenes. This technique, somehow equivalent to an adaptive filtering technique, has been studied and tailored to fit with our expectations. The EMD has shown a great ability to remove efficiently the random fluctuating background intensity and to evaluate the modulation intensity. The Hilbert tranform (HT) is the natural quadrature operator. Its use to build an analytical signal from the so-detrended SI signal, for subsequent phase computation, has been studied and assessed. Other phase extraction techniques have been considered as well for comparison purposes. Finally, our answer to the decorrelation-induced phase error relies on the well-known result that the higher the pixel modulation intensity, the lower the random phase error. We took benefit from this result – not only linked to basic SNR considerations, but more specifically to the intrinsic phase structure of speckle fields – with a novel approach. The regions within the pixel signal history classified as unreliable because under-modulated, are purely and simply discarded. An interpolation step with the Delaunay triangulation is carried out with the so-obtained non-uniformly sampled phase maps to recover a smooth phase which relies on the most reliable available data. Our schemes have been tested and discussed with simulated and experimental SI signals. We eventually have developed a versatile, accurate and efficient phase extraction procedure, perfectly able to tackle the challenge of dynamic behaviors characterization, even for displacements and/or deformations beyond the classical limit of the correlation dimensions.