This paper presents a non-linear approacb tor acoustical pre-processing based on Veetor Quantization. Tbe idea is to transform feature vectors extracted from signais of one quality to feature vectors of another quality. Our method is applied to two particular cases: speech recognition at multiple sampling rates, and speech recognition in noisy environment. Such a method, which could be applied to other adaptation problems, allows very acceptable correspondence between two considered feature spaces. Tbus, a genene ASR system trained on 16kHz signais is able to recognize lower sampling rate signals without any adaptation of its acoustic modeis. In the same way, our method is applied to ASR in noisy environment and its performance is better than conventional MLLR adaptation when large amount of adaptation data is provided.