To provide a voice recognition device in which high-speed noise adaptation can be performed to a voice model.

When a noise adaptive voice model is formed by a noise model and a voice model by means of PMC (noise adaptation), PMC conversion of the whole distribution of all models is not performed as in the past, but the distribution of the voice model is grouped or clustered to form a plurality of typical distributions in a noise adaptive part 204, each representative distribution and the positional relation of its group internal distribution are stored, and only the typical distribution is PMC converted. The noise adaptive voice model is formed on the basis of the PMC-converted typical distribution and the stored positional relation.