Abstract: An object of the present invention is to facilitate dealing with noisy speech with varying SNR and save calculation costs by generating a speech model with a single-tree-structure and using the model for speech recognition. Every piece of noise data stored in a noise database is used under every SNR condition to calculate the distance between all noise models with the SNR conditions and the noise-added speech is clustered. Based on the result of the clustering, a single-tree-structure model space into which the noise and SNR are integrated is generated (steps S1 to S5). At a noise extraction step (step S6), inputted noisy speech to be recognized is analyzed to extract a feature parameter string and the likelihoods of HMMs are compared one another to select an optimum model from the tree-structure noisy speech model space (step S7). Linear transformation is applied to the selected noisy speech model space so that the likelihood is maximized (step S8).