This paper addresses the problem of the mismatch between a silencemodel and background noises which often occurs in a telephone speechrecognition system (SRS) application. At first, the use of parallelmodel combination (PMC) methods is studied with the respect to thisapplication. Secondly, the effective adaptation of a silence model tovarious background noises is confirmed. Finally, an original methodcombining log-add PMC with a noise power spectral density estimationbased on minimum statistics is proposed. The performed tests prove thebenefit of the suggested method to the speech recognition results thatis caused by the stability of speech vector selection under theinfluence of various background noises. The advantages can be seen inno extra voice activity detector and in a relatively low computationalload.