Automatic pitch marking for speech transformations via TD-PSOLA

This paper describes an automatic pitch marking method which can be used in the context of modifying speech signals with TD-PSOLA (Time Domain Pitch Synchronous Overlap-Add). Unlike other approaches where glottal closure instants arc searched directly in the speech signal, our approach exploits results of a pitch extraction algorithm. The principle is to optimize the propagation of pitch marks from one pitch period to the following by means of dynamic programming. The first step consists of extracting extrema on regularly spaced segments, the size of which is the smallest pitch period in the signal under investigation. Then, an optimal subset of extrema, which represent pitch marks, is found by a dynamic programming algorithm which derives from a smoothing algorithm proposed by Ney. Pitch marks obtained through this post-synchronization algorithm allow very good speech signal transformations. Furthermore, this algorithm has the advantage that it is independent from the pitch determination step which allows it to be easily combined with any algorithm for pitch extraction.