Spoken language estimating device, method, and program

PROBLEM TO BE SOLVED: To estimate the kind of a language indicated by an input speech signal without converting into text-level language expressions, also without requiring previous knowledge.SOLUTION: In a spoken language estimating device, a phoneme expression calculation section 13 decomposes a mel spectrum extracted from a speech signal for learning with NMF (Non-negative Matrix Factorization) to obtain a phoneme expression H and a compound ratio U for each kind of language, and the phoneme expression H is stored in a phoneme expression storage section 14 for each kind of language. When a speech signal for estimating is input, a feature information extraction section 12 extracts a mel spectrum, and a phoneme compound ratio calculation section 15 calculates a compound ratio U on the basis of the extracted mel spectrum and the phoneme expressions H stored in the phoneme expression storage section 14, for each kind of language. A language similarity estimation section 16 calculates the product of the calculated compound ratio U and the phoneme expressions H stored in the phoneme expression storage section 14, for each kind of language, to estimate the kind of language indicated by the speech signal for estimating on the basis of similarity between the product and the mel spectrum extracted from the speech signal for estimating.