This paper presents a real-time kepstrum analysis algorithm which gives improvement of the SNR of speech contaminated with additive nonstationary noise. The kepstrum method (complex cepstrum) is applied to the front-end of a conventional speech enhancement method. The test, applied to the modified Griffith and Jim adaptive beamformer, shows an improvement of 6.55dB during speech and nonstationary noise periods. Experiments on a 20cm broadside microphone configuration are implemented in realtime in a real environment (indoor office)