In this paper, we describe a combination of normalization techniques for removing the effects of additive noise, convolutional noise and speech level variations on the speech representation that is produced by an auditory model. In [1, 2], we introduced a noise magnitude subtraction technique (NMS) that removes the effects of additive noise. In [2], we presented a loudness mean normalization technique (LMN) that compensates for speech level variations and convolutional noise. In this paper, an improved technique, called acoustic model controlled loudness mean normalization (ALMN) is introduced. It is shown that ALMN, when tested on convolutional noise, consistently outperforms the standard LMN, especially if less than 2 seconds of speech are used to estimate the noise characteristics. In fact, the ALMN performance does not degrade substantially until less than 1 second of speech is used for estimation. It is shown that the ALMN technique, when used in combination with NMS, offers a highly efficient joint compensation for multiple noise-types. Again, the combination ALMN+NMS consistently outperforms the combination LMN+NMS.