Granted Patent B2

US 12,354,608 · App. 18/321,353 · Granted Jul 8, 2025

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

Assignee: Pindrop Security, Inc.

G10L17/20G10L17/02G10L17/04G10L17/18G10L19/028

Abstract

Claims (38)

1. A computer-implemented method comprising:

obtaining, by a computer, one or more enrolling speech signals for a registered speaker during speaker registration;

generating, by the computer, one or more degraded enrolling signals by applying an acoustic channel simulator on the one or more enrolling speech signals;

extracting, by the computer, an enrolled voiceprint for the registered speaker by applying a trained neural network on the one or more degraded enrolling signals and the one or more enrolling speech signals;

obtaining, by the computer, a test speech signal for a test speaker during speaker recognition;

extracting, by the computer, channel-compensated features from the test speech signal for the test speaker by applying the trained neural network on the test speech signal; and

identifying, by the computer, the test speaker as the registered speaker in response to determining that a loss between the test voiceprint and the enrolled voiceprint satisfies a distance threshold.

2. The method according to claim 1 , wherein the acoustic channel simulator generates a degraded enrolling speech signal having a type of degradation added to an enrolling speech signal from the registered speaker, and wherein the type of degradation includes one of environmental noise, reverberation, an audio acquisition device characteristic, or an audio channel transcoding artifact.

3. The method according to claim 1 , wherein generating a degraded enrolling speech signal includes:

selecting, by the computer, an environmental noise type as a type of degradation from a plurality of environmental noise types stored in a noise profile database; and

simulating, by the computer, the environmental noise type in the enrolling speech signal applied by the acoustic channel simulator on an enrolling speech signal.

4. The method according to claim 1 , wherein generating the degraded enrolling signal includes simulating, by the computer, a set of one or more audio acquisition device characteristics according to an audio acquisition device profile applied by the acoustic channel simulator to an enrolling speech signal.

5. The method according to claim 4 , wherein each audio acquisition device profile includes at least one of a frequency characteristic, an amplitude characteristic, a filtering characteristic, an electrical noise characteristic, or a physical noise characteristic.

6. The method according to claim 1 , wherein generating the degraded enrolling speech signal includes:

selecting, by the computer, a transcoding profile from a plurality of transcoding profiles stored in a transcoding profile database; and

simulating, by the computer, a set of audio channel transcoding characteristics according to the transcoding profile applied by the acoustic channel simulator to an enrolling speech signal.

7. The method according to claim 1 , wherein generating the degraded enrolling speech signal includes simulating, by the computer, a reverberation according to a direct-to-reverberation ratio (DRR) applied by the acoustic channel simulator to the enrolling speech signal.

8. The method according to claim 1 , further comprising training, by the computer, a neural network to minimize a loss between a set of channel-compensated training features extracted from a training signal by applying the neural network and a set of handcrafted training features extracted from corresponding clean training speech signals, thereby training the trained neural network.

9. The method according to claim 8 , further comprising applying, by the computer, the acoustic channel simulator on a clean training signal to generate the training signal having a type of degradation corresponding to a channel-compensated training feature for training the trained neural network.

10. The method according to claim 8 , wherein a handcrafted training features includes at least one of: Mel-frequency cepstrum coefficients (MFCCs), low-frequency cepstrum coefficients (LFCCs), perceptual linear prediction (PLP) coefficients, linear or Mel filter banks, or glottal features.

11. A computer-implemented method comprising:

obtaining, by a computer, one or more training speech signals containing corresponding training utterances;

for a training speech signal of the one or more training speech signals,

generating, by the computer, a degraded training speech signal having a type of degradation by applying a noise simulator on the training speech signal;

extracting, by the computer, a set of training low-level features for the degraded speech signal by applying a neural network on the degraded speech signal;

extracting, by the computer, a set of training handcrafted features for the speech signal by applying a signal analyzer on the training speech signal; and

updating, by the computer, one or more connection values of the neural network in response to determining that a loss result fails to satisfy a training loss threshold, the loss result based upon the set of training low-level features and the set of training handcrafted features.

12. The method according to claim 11 , further comprising generating, by the computer, the loss result according to a loss function using the set of training low-level features and the set of training handcrafted features.

13. The method according to claim 11 , wherein, for each type of degradation of a plurality of types of degradation, the computer updates the one or more connection values of the neural network according to a loss function to minimize the loss result between the set of training low-level features and the set of training handcrafted features.

14. The method according to claim 11 , wherein, for each training speech of the one or more training speech signals, the computer updates the one or more connection values of the neural network according to a loss function to minimize the loss result between the set of training low-level features and the set of training handcrafted features.

15. The method according to claim 11 , further comprising:

determining, by the computer, that a next loss result for a next training signal satisfies the training loss threshold; and

storing, by the computer, a trained neural network from the neural network in response to determining that the next loss result satisfies the training loss threshold.

16. The method according to claim 11 , wherein generating the degraded training speech signal having the type of degradation includes simulating, by the computer, an environmental noise being applied to the training speech signal according to a type of environmental noise.

17. The method according to claim 11 , wherein generating the degraded training speech signal having the type of degradation includes simulating, by the computer, a reverberation being applied to the training speech signal according to a direct-to-reverberation ratio (DRR).

18. The method according to claim 11 , wherein generating the degraded training speech signal having the type of degradation includes simulating, by the computer, one or more audio device characteristics being applied to the training speech signal according to an audio acquisition device profile.

19. The method according to claim 11 , wherein generating the degraded training speech signal having the type of degradation includes simulating, by the computer, a set of audio channel transcoding characteristics being applied to the training speech signal according to a transcoding profile.

20. The method according to claim 11 , wherein the low-level features include at least one of: Mel-frequency cepstrum coefficients (MFCCs), low-frequency cepstrum coefficients (LFCCs), perceptual linear prediction (PLP) coefficients, linear or Mel filter banks, or glottal features.

Assignments (4)

Continuity (6)

References Cited (339)