We describe a novel algorithm for formant ex- traction and tracking to be used in the context of acoustic-phonetic decoding. The construction of formant tracks involves both image processing techniques and geometric reasoning, which constitutes the originality and efficiency of the method. The algorithm which works either on LPC or on cepstral data, comprises two distinct steps: - first, formant segments are extracted from the signal by means of local processing, - secondly geometric reasoning involving acoustic-phonetic knowledge is applied, in order to connect meaningful elementary tracks consistently and eliminate irrelevant segments.