Normalisation of a verification measure in a device for the speaker verification

摘要

A device for speech recognition generates parameters (m) of an acceptance voice model (λ) relative to a voice segment (mp) delivered by an authorized speaker and of a voice model of the rejects (?) during a learning phase. The normalized by the parameters of standardization ( λ, σλ), a verification score of the speaker (sv) depending on the ratio of likelihood between a voice segment to be tested (x) and the models of acceptance and rejects. The speaker accesses a service application (ap), if the normalized score (sn) is greater than a threshold (th1). According to the invention, a module (v6) updates the parameters of standardization ( λ, τλ) as a function of the verification score (sv) to each test of the voice segment only if the normalized score (sn) is greater than a second threshold (th2).