To perform speech recognition in a noisy environ- ment, it is important to minimize the influences of both additive noise and the Lombard effect. Data analysis clearly showed a trend of formant frequency shift in some specific frequency bands due to the Lombard effect. This paper reports a method of modifying the LPC cepstrum to compensate for the formant frequency shift, which had usually been left uncorrected. The compensated value can be obtained by multiplication of the formant frequency shift value and the partial differential of the LPC cepstrum with respect to the formant-frequency. To validate the effect of the compensation method proposed herein, the following items were studied. (1) Comparison to the usual method (Compensation of spectral tilt). (2) Comparison to the weighted cepstrum.