Voice of a user is inputted to a speech recognition section until a start of a no-voice domain from depression of a talk-switch. LPC cepstrum coefficients are calculated from the voice in an LPC analysis section and a cepstrum calculation section, and then temporarily stored in a parameter backward output section. A series of the LPC cepstrum coefficients is re-arranged to the series in which the time axis is inverted and then outputted to a collating section. The collating section calculates a degree of similarity between the LPC cepstrum coefficients and a recognition dictionary of a backward tree-structure stored in a standard pattern section through a backward collating.