Formant tracking is one of the most important issues in speech recognition. Nevertheless, most algorithms are based on local methods and take into account global formant properties with difficulty. We therefore propose a formant tracking algorithm which provides a global point of view on formant tracking. The salient idea is to combine local tracking to generate elementary formant tracking hypotheses and an active method to regularize global formant trajectories in the following way: the formant trajectory is the closest curve to that of the hypothesis maximizing energy incorporated by the formant and which is sufficiently smooth. The main qualities of this algorithm are robustness and ability to detect accurate and regular formant trajectories.