Verification score normalization in a speaker voice recognition device
AssigneeFrance Telecom
InventorDelphine Charlet
Life of this patent
Estimating the term…
Abstract
During a learning phase, a speech recognition device generates parameters of an acceptance voice model relating to a voice segment spoken by an authorized speaker and a rejection voice model. It uses normalization parameters to normalize a speaker verification score depending on the likelihood ratio of a voice segment to be tested and the acceptance model and rejection model. The speaker obtains access to a service application only if the normalized score is above a threshold. According to the invention, a module updates the normalization parameters as a function of the verification score on each voice segment test only if the normalized score is above a second threshold.
Claims10
1- a device for automatically recognizing the voice of a speaker authorized to access an application, said device comprising means for generating beforehand, during a learning phase, parameters of an acceptance voice model relative to a voice segment spoken by said authorized speaker and parameters of a rejection voice model, means for normalizing by means of normalization parameters a speaker verification score depending on the likelihood ratio between a voice segment to be tested and said acceptance model and rejection model thereby driving a normalized verification score, and means for comparing said normalized verification score to a first threshold in order to authorize access to the application by the speaker who spoke said voice segment to be tested only if the normalized verification score is at least as high as the first threshold, and means for updating at least one of said normalization parameters as a function of a preceding value of said one normalization parameter and the speaker verification score on each voice segment test only if the normalized verification score is at least equal to a second threshold that is at least equal to said first threshold.
2- A device according to claim 1, wherein said normalization parameter updated is representative of a statistical mean value of the speaker verification score.
3- A device according to claim 2, wherein said statistical mean value {tilde over (μ)}λ of the speaker verification score SV is updated in accordance with the following relationship:
- {tilde over (μ)}λ≡(1−τμ){tilde over (μ)}λ+τμ·SV
- in which τμ is a predetermined adaptation factor.
4- A device according to claim 3, wherein said predetermined adaptation factor τμ varies as a function of the number of normalization parameter updates.
5- A device according to claim 1, wherein the parameter updated is representative of the standard deviation of said speaker verification score.
6- A device according to claim 5, wherein said standard deviation {tilde over (σ)}λ of the speaker verification score SV is updated in accordance with the following relationship:
- {tilde over (σ)}λ≡{square root}{square root over ((1−τσ){tilde over (σ)})}λ2+τσ(SV−{tilde over (μ)}λ)2
- in which τσ is a predetermined adaptation factor.
7- A device according to claim 6, wherein said predetermined adaptation factor τσ varies as a function of the number of normalization parameter updates.
8- A device according to claim 1, comprising means for updating at least one of said parameters of said acceptance voice model as a function of a preceding value of said model parameter only if the normalized verification score is at least equal to said second threshold.
9- A device according to claim 8, wherein said model parameter m is updated in accordance with the following equation:
- 5 m = N AP m AP + N adapt m adapt N AP + N adapt
- in which mAP and NAP respectively denote a mean value of Gaussian distribution of probability density of said model parameter m during said learning phase and the number of frames in voice segments used to estimate mean values of Gaussian distributions relative to said acceptance model and rejection model, madapt denotes a mean value of Gaussian distribution of probability density of said model parameter m determined during the update that has just been effected, and Nadapt denotes the number of frames used to estimate a mean value of the Gaussian distribution of said model parameter m for said update that has just been effected.
10- A device according to claim 1, wherein said normalized verification score SN is determined as a function of said speaker verification score SV and two updated normalization parameters {tilde over (μ)}λ and {tilde over (σ)}λ, in accordance with the following equation:
- 6 S N = S V - μ ~ λ σ ~ λ ,
- in which said parameters {tilde over (μ)}λ and {tilde over (σ)}λ are respectively the statistical mean value and the standard deviation of said speaker verification score.