A speech codec includes a drive sound source generating circuit in which, in a case of a voiced sound speech, a pulse pattern signal corresponding to a pitch-scale, drive sound source signals stored within a newest predetermined past time period and a noise signal are multiplied by predetermined gains, respectively, and then, added to each other so as to generate a drive voiced sound source and, in a case of an unvoiced sound, drive sound source signals stored within a newest predetermined past time period and a noise signal are multiplied by predetermined gains, respectively, and then, added to each other so as to generate a drive unvoiced sound source.