PURPOSE: To perform variable speed control over high-speed voicing by preparing plural prediction gain threshold values for frame parameters corresponding to voicing speeds by learning, and changing threshold values selected among them and controlling the quantity of frames to be thinned out.

CONSTITUTION: A prediction gain calculator 3 calculates a prediction gain from spectrum information sent, frame by frame, from a voice memory 2. This prediction gain is compared by a decision device 4 with one threshold value selected from a threshold table 14 by a threshold value control circuit 13 according to a speed parameter sent from a host CPU 12. When the prediction gain is less than the threshold value, a synthesizing filter B is interrupted under the control of a buffer control circuit 11 to thin out frames.