Abstract: A plurality of pitch period transition paths are extracted by a pitch tracking over a frame, and a path of a minimum average prediction gain over the frame is selected from the extracted paths. A subsequent preliminary pitch selection may be executed in a sub-frame processing to select a plurality of candidates from the neighborhood of the pitch of the transition path selected for each sub-frame by using the inner product of the input speech signal and each codevector. Finally, a pitch period having a minimum waveform distortion is selected for each sub-frame.