In a speech processor such as a speech recogniser, the problem of distortion of extracted features caused by adaption of the input automatic gain control (AGC) during feature extraction is solved by storing the AGC's gain coefficient along with the energy level of each extracted feature. At the end of the sampling period the stored gain coefficients are set equal to the minimum stored coefficient and the associated energy levels adjusted accordingly. The AGC circuit may comprise a digitally switched attenuator under the control of a microprocessor performing the speech recognition.
