An input speech signal is splitted into a time section and the splitted signal is generated as a speech vector and LPC coefficient sets are developed by linear prediction analysis for every time section of the speech vector. The speech vector is weighted based on the developed LPC coefficient sets, then a plurality of the weighted speech vectors are connected and the connected speech vector having a predetermined frame length is generated. An excitation codevector whose weighted synthesized signal is the most similar to the weighted speech vector is determined among from a plurality of excitation vectors each having the frame length which are previously stored as a sound source. A plurality of adaptive codevectors each having the frame length and obtained by cutting out a sound source signal produced from the determined excitation vectors at predetermined timing points are stored in an adaptive codebook (145). An adaptive codevector whose weighted synthesized signal is the most similar to the weighted speech vector is determined among from the plurality of adaptive codevectors.