PURPOSE: To reduce the occurrence of erroneous detection of a voice section without lowering a processing speed by detecting a section where a voice is present by means of an output through a preemphasis to emphasize the component of a frequency approximately equal to the characteristic frequency region frequency of a vowel having a low voice power.

CONSTITUTION: A voice input part 1 collects voices and converts the voices into an electric signal xt, and a preemphasis 2 emphasizes the component of a frequency approximately equal to the characteristic frequency region frequency of a voice to increase a voice of a vowel having a low voice power. A threshold calculating part 3 selectively calculates a threshold and calculates a threshold Th by means of an output signal Pr from the preemphasis 2 at a section where no voice is present. A section detecting part 4 detects a section by means of signals Pr and Tg, an output signal and an input signal therefrom are inputted to a voice input part 5 to recognize a voice.