Background noise is one of the biggest problem for speech recognition systems in real environments. In order to achieve high recognition performance for corrupted speech, we proposed a new construction method of HMMs dealing with various kinds of background noise. At first, each HMM dealing with a single noise is trained for each background noise, and then all Gaussian components of those HMMs are combined into a "multi-mixture HMM". From the experimental results, the multi-mixture HMM gave the highest recognition performance for any kind of noise and any variation of SNR.

Although the multi-mixture HMMs has high performance, it has a huge number of Gaussian components that makes the speech recognition slower. In order to solve the problem, we also proposed a reduction method of Gaussian components. It can decrease the number of Gaussian components with slight deterioration of recognition performance.