Methods and systems for automatic discovery of fraudulent calls using speaker recognition

1. A computer-implemented method for determining voices associated with undesirable activity, comprising:

receiving a plurality of audio recordings, the plurality of audio recordings comprising voices associated with undesirable activity;

determining a plurality of audio components of each of the plurality of audio recordings;

generating a multi-dimensional vector of audio components, from the plurality of audio components, for each of the plurality of audio recordings to generate a plurality of multi-dimensional vectors of audio components;

comparing audio components between the plurality of multi-dimensional vectors of audio components to determine a plurality of clusters of multi-dimensional vectors, each cluster of the plurality of clusters comprising two or more of the plurality of multi-dimensional vectors of audio components, wherein each cluster of the plurality of clusters corresponds to a blacklisted voice;

determining at least one erroneous cluster in the plurality of clusters by filtering the plurality of clusters, according to one or more predetermined thresholds, based on at least one of: cluster size, cluster coherence, cluster member metadata, a distribution of similarity scores, or a proportion of audio recordings associated with each of the plurality of clusters confirmed to be associated with undesirable activity;

removing the erroneous cluster from the plurality of clusters;

receiving an audio recording or audio stream; and

determining whether the audio recording or audio stream is associated with a voice associated with undesirable activity based on a comparison to the plurality of clusters.

2. The method of claim 1 , further comprising:

performing a nearest neighbor search on each of the plurality of multi-dimensional vectors of audio components to determine a nearest neighbor metric; and

discarding one or more of the plurality of audio recordings in response to determining that the one or more of the plurality of audio recordings has a nearest neighbor metric smaller than a predetermined nearest neighbor metric.

3. The method of claim 2 , wherein comparing audio components between the plurality of multi-dimensional vectors of audio components comprises determining a minimum similarity between each of the plurality of multi-dimensional vectors of audio components and/or determining an average similarity between each of the plurality of multi-dimensional vectors of audio components.

4. The method of claim 1 , wherein determining whether an audio recording or audio stream is associated with a voice associated with undesirable activity comprises:

receiving an audio recording, the audio recording comprising a voice;

determining a plurality of audio components of the audio recording;

generating a multi-dimensional vector of audio components of the audio recording;

determining whether the multi-dimensional vector of audio components matches with one of the plurality of clusters of multi-dimensional vectors; and

in response to determining that the multi-dimensional vector of audio components clusters with one of the plurality of clusters of the multi-dimensional vectors, flagging the audio recording as being associated with a voice associated with undesirable activity.

5. The method of claim 1 , wherein the plurality of clusters of multi-dimensional vectors is determined without human assistance.

6. The method of claim 1 , wherein each multi-dimensional vector of audio components of the plurality of multi-dimensional vectors of audio components comprises a fixed number of dimensions.

7. The method of claim 1 , wherein comparing audio components between the plurality of multi-dimensional vectors of audio components further comprises performing agglomerative hierarchical clustering across the plurality of audio recordings.

8. The method of claim 1 , further comprising:

upon determining that the audio recording or audio stream is associated with a voice associated with undesirable activity, determining a genuine account based on the audio recording or audio stream;

determining the verified owner of the genuine account; and

generating an alert to the verified owner of the genuine account.

9. The method of claim 1 , wherein comparing audio components between the plurality of multi-dimensional vectors of audio components further comprises concatenating metadata, associated with the plurality of audio recordings, with the plurality of the plurality of multi-dimensional vectors.

10. The method of claim 1 , wherein the comparison to the plurality of clusters comprises determining a cosine similarity or log-likelihood ratio.

11. The method of claim 1 , further comprising:

storing the plurality of clusters of multi-dimensional vectors in an undesirable voice data store.

12. A computer-implemented method for determining voices associated with undesirable activity, comprising:

receiving an audio recording, the audio recording being associated with a voice associated with undesirable activity;

determining a plurality of audio components of the audio recording;

generating a multi-dimensional vector of audio components of the audio recording;

determining whether the multi-dimensional vector of audio components matches with one of a plurality of clusters of multi-dimensional vectors, the plurality of clusters of multi-dimensional vectors being associated with a plurality of audio recordings, the plurality of audio recordings comprising voices associated with undesirable activity, the plurality of clusters of multi-dimensional vectors being automatically generated based on a plurality of audio recordings without human intervention;

determining at least one erroneous cluster in the plurality of clusters of multi-dimensional vectors by filtering the plurality of clusters of multi-dimensional vectors, according to one or more predetermined thresholds, based on at least one of: cluster size, cluster coherence, cluster member metadata, a distribution of similarity scores, or a proportion of audio recordings associated with each of the plurality of clusters of multi-dimensional vectors confirmed to be associated with undesirable activity;

removing the erroneous cluster from the plurality of clusters of multi-dimensional vectors; and

in response to determining that the multi-dimensional vector of audio components matches one of the plurality of clusters of the multi-dimensional vectors, flagging the audio recording as potentially undesirable.

13. The method of claim 12 , wherein determining that the multi-dimensional vector of audio components matches one of the plurality of clusters comprises determining a cosine similarity or log-likelihood ratio.

14. The method of claim 12 , wherein each of the plurality of audio recordings contains a single distinct voice.

15. The method of claim 12 , wherein the multi-dimensional vector of audio components comprises a fixed number of dimensions.

16. The method of claim 12 , wherein determining whether the multi-dimensional vector of audio components matches with one of a plurality of clusters of multi-dimensional vectors is determined without human assistance.

17. The method of claim 12 , further comprising:

upon determining that the audio recording or audio stream is associated with a potentially undesirable voice, determining a genuine account based on the audio recording or audio stream;

determining the verified owner of the genuine account; and

generating an alert to the verified owner of the genuine account.

18. The method of claim 12 , wherein the multi-dimensional vector of audio components is stored in a voice data store.

19. A computer system for determining potentially undesirable voices, the computer system comprising:

a memory storing instructions; and

one or more processors configured to execute the instructions to perform operations including:

receiving a plurality of audio recordings, the plurality of audio recordings comprising voices associated with undesirable activity;

determining a plurality of audio components of each of the plurality of audio recordings;

generating a multi-dimensional vector of audio components, from the plurality of audio components, for each of the plurality of audio recordings to generate a plurality of multi-dimensional vectors of audio components;

comparing audio components between the plurality of multi-dimensional vectors of audio components to determine a plurality of clusters of multi-dimensional vectors, each cluster of the plurality of clusters comprising two or more of the plurality of multi-dimensional vectors of audio components, wherein each cluster of the plurality of clusters corresponds to a blacklisted voice;

determining at least one erroneous cluster in the plurality of clusters by filtering the plurality of clusters, according to one or more predetermined thresholds, based on at least one of: cluster size, cluster coherence, cluster member metadata, a distribution of similarity scores, or a proportion of audio recordings associated with each of the plurality of clusters confirmed to be associated with undesirable activity;

removing the erroneous cluster from the plurality of clusters;

receiving an audio recording or audio stream; and

determining whether the audio recording or audio stream is associated with a voice associated with undesirable activity based on a comparison to the plurality of clusters.