Abstract—To improve the robustness of distributed speech front-ends in mobile communication we introduce, in this paper, a new set of feature vector which is estimated through three steps. First, the Mel-Line Spectral Frequencies (MLSFs) coefficients are combined with conventional MFCCs, after extracted from a denoised acoustic frame using the wiener filter. Also, we optimize the stream weights of multi-stream HMMs by deploying a discriminative approach. Finally, these features are adequately transformed and reduced in a multi-stream scheme using Karhunen-Loeve Transform (KLT). Recognition experiments on the Aurora 2 connected digits database reveal that the proposed front-end leads to a significant improvement in speech recognition accuracy for highly noisy GSM.