Speech recognition is improved using reference pattern templates which have an added noise signal (noise floor) to avoid LPC high-gain synthesizer instability at low signal levels. Also, input signal frames have a length one-half that of reference frames whereby dynamic time warp computation steps are cut almost in half.