Granted Patent B2

US 12,512,101 · App. 17/963,091 · Granted Dec 30, 2025

Inventors: Elie Khoury (Atlanta, GA); Matthew Garland (Atlanta, GA)

Assignee: Pindrop Security, Inc.

G10L17/08G06N3/04G06N3/08G10L15/16G10L17/02G10L17/04G10L17/18G10L17/22

Abstract

Claims (44)

1 . A computer-implemented method comprising:

obtaining, by the computer, one or more enrollment speech samples associated with an enrolled speaker;

executing, by the computer, a neural network on the one or more enrollment speech samples to generate an enrolled voiceprint for the enrolled speaker, wherein the neural network is trained according to positive speech samples and negative speech samples;

obtaining, by the computer, a recognition speech sample associated with a real-time speaker;

executing, by the computer, the neural network on the recognition speech sample to generate a real-time voiceprint for the real-time speaker;

generating, by the computer, a cosine distance between the enrolled voiceprint and the real-time voiceprint indicating an amount of similarity between the enrolled speaker and the real-time speaker; and

authenticating, by the computer, the real-time speaker as the enrolled speaker of a closed set of blacklisted speakers according to the cosine distance.

2 . The method according to claim 1 , further comprising identifying, by the computer, the real-time speaker is the enrolled speaker, in response to determining that the cosine distance satisfies a recognition threshold.

3 . The method according to claim 2 , wherein the enrolled voiceprint of the enrolled speaker is associated with the closed set of blacklisted speakers, and wherein the computer blocks a call of the real-time speaker.

4 . The method according to claim 1 , wherein the neural network is a portion of a triplet neural architecture.

5 . The method according to claim 1 , wherein the neural network is trained by reducing a loss function including a training cosine distance for dual sets of the positive speech samples and a cohort set of the negative speech samples.

6 . The method according to claim 1 , further comprising:

feeding, by the computer, a set of positive training speech samples into one or more feed-forward neural networks including the neural network to generate one or more corresponding embedding vectors;

calculating, by the computer, a loss function based upon the one or more embedding vectors; and

back-propagating, by the computer, the loss function to modify one or more connection weights in each of the feed-forward neural networks.

7 . The method according to claim 6 , wherein the loss function is based upon a positive cosine distance corresponding to a degree of similarity between a first embedding vector and a second embedding vector.

8 . The method according to claim 6 , further comprising initializing, by the computer, at least one feed-forward neural network with random connection weights.

9 . The method according to claim 1 , further comprising pre-processing, by the computer, the speech sample prior to executing the neural network, wherein the speech sample is at least one of an enrollment speech sample or real-time speech sample.

10 . The method according to claim 9 , further comprising:

segmenting, by the computer, the speech sample into windows of a predetermined duration with a predetermined window shift; and

extracting, by the computer, a set of features to be fed into the neural network from each window.

11 . A system comprising:

a non-transitory storage medium storing a plurality of computer program instructions; and

a computer comprising a processor electrically coupled to the non-transitory storage medium and configured to execute the plurality of computer program instructions to:

obtain one or more enrollment speech samples associated with an enrolled speaker;

execute a neural network on the one or more enrollment speech samples to generate an enrolled voiceprint for the enrolled speaker, wherein the neural network is trained according to positive speech samples and negative speech samples;

obtain a recognition speech sample associated with a real-time speaker;

execute the neural network on the recognition speech sample to generate a real-time voiceprint for the real-time speaker;

generate a cosine distance between the enrolled voiceprint and the real-time voiceprint indicating an amount of similarity between the enrolled speaker and the real-time speaker; and

authenticate the real-time speaker as the enrolled speaker of a closed set of blacklisted speakers according to the cosine distance.

12 . The system according to claim 11 , wherein the computer is further configured to identify the real-time speaker is the enrolled speaker, in response to determining that the cosine distance satisfies a recognition threshold.

13 . The system according to claim 12 , wherein the enrolled voiceprint of the enrolled speaker is associated with a closed set of blacklisted speakers, and wherein the computer blocks a call of the real-time speaker.

14 . The system according to claim 11 , wherein the neural network is a portion of a triplet neural architecture.

15 . The system according to claim 11 , wherein the neural network is trained by reducing a loss function including a training cosine distance for dual sets of the positive speech samples and a cohort set of the negative speech samples.

16 . The system according to claim 11 , wherein the computer is further configured to:

feed a set of positive training speech samples into one or more feed-forward neural networks including the neural network to generate one or more corresponding embedding vectors;

calculate a loss function based upon the one or more embedding vectors; and

back-propagate the loss function to modify one or more connection weights in each of the feed-forward neural networks.

17 . The system according to claim 16 , wherein the loss function is based upon a positive cosine distance corresponding to a degree of similarity between a first embedding vector and a second embedding vector.

18 . The system according to claim 16 , wherein the computer is further configured to initialize at least one feed-forward neural network with random connection weights.

19 . The system according to claim 11 , wherein the computer is further configured to pre-process the speech sample prior to executing the neural network, wherein the speech sample is at least one of an enrollment speech sample or real-time speech sample.

20 . The system according to claim 19 , wherein the computer is further configured to:

segment the speech sample into windows of a predetermined duration with a predetermined window shift; and

extract a set of features to be fed into the neural network from each window.

Assignments (4)

Continuity (4)

References Cited (346)