Highly accurate personal identification systems are required in many different recognition situations. In this study, the Mel-frequency cepstrum coefficient was used to extract the features of speakers. The aim of this study was to identify different speeches in different noisy environments. A maximum likelihood estimation method based on noise probability was proposed to enhance the recognition effects of the Gaussian mixture model of different speeches from mixed noise speech signals. Experimental results indicated that the method had high recognition results under various noise conditions. The recognition results of the proposed method in different noise environments were superior to those of a method using only one type of noise for modeling. Experimental results obtained from some unspecified speakers showed that three different languages (Mandarin, English, and Cantonese) were effectively identified.