Unsupervised keyword spotting and word discovery for fraud analytics

Patent metadata

Field	Value
Publication number	US-2022301554-A1
Application number	US-202217833674-A
Country	US
Kind code	A1
Filing date	Jun 6, 2022
Priority date	Jan 28, 2019
Publication date	Sep 22, 2022
Grant date	—

Abstract

Embodiments described herein provide for a computer that detects one or more keywords of interest using acoustic features, to detect or query commonalities across multiple fraud calls. Embodiments described herein may implement unsupervised keyword spotting (UKWS) or unsupervised word discovery (UWD) in order to identify commonalities across a set of calls, where both UKWS and UWD employ Gaussian Mixture Models (GMM) and one or more dynamic time-warping algorithms. A user may indicate a training exemplar or occurrence of call-specific information, referred to herein as “a named entity,” such as a person's name, an account number, account balance, or order number. The computer may perform a redaction process that computationally nullifies the import of the named entity in the modeling processes described herein.

First claim

What is claimed is:

1 . A computer-implemented method comprising:
generating, by a computer, a plurality of segments of a plurality of audio signals, including one or more candidate segments of a first audio signal and one or more query segments of a second audio signal; extracting, by the computer, a plurality of features for each segment of the plurality of segments; determining, by the computer, for a candidate segment of the first audio signal a plurality of similarity scores corresponding to the plurality of query segments of the second audio signal, each similarity score indicating a similarity of the features of the candidate segment with respect to the features of a corresponding query segment of the second audio signal; identifying, by the computer, a discovered keyword included in the candidate segment, in response to determining that at least one similarity score of the candidate segment satisfies a pairwise matching threshold; and updating, by the computer, a voice detection model associated with the plurality of audio signals to nullify a probability of a portion of an audio signal having the discovered keyword.

2 . The method according to claim 1 , wherein identifying the discovered keyword in the candidate segment includes determining, by the computer, that the similarity score for the candidate segment satisfies the pairwise matching threshold.

3 . The method according to claim 1 , wherein generating the plurality of segments of the plurality of audio signals includes:
parsing, by the computer, the candidate segment of the first audio signal into a plurality of candidate paths, each candidate path comprising a fixed portion of the candidate segment; and parsing, by the computer, the query segment of the second audio signal into a plurality of test paths, each test path comprising the fixed portion of the query segment.

4 . The method according to claim 3 , wherein the computer determines the similarity score for the candidate segment based upon comparing the plurality of candidate paths of the candidate segment and the plurality of test paths in the query segment, and
wherein the computer identifies the discovered keyword in the candidate segment based upon determining that the similarity score for the candidate segment satisfies the pairwise-match threshold score.

5 . The method according to claim 4 , wherein identifying the discovered keyword in the candidate segment includes determining, by the computer, that a threshold number of candidate paths of the candidate segment satisfy a second threshold score.

6 . The method according to claim 1 , further comprising:
identifying, by the computer, a pairwise match including the candidate segment of the first audio signal and the query segment of the second audio signal having a similarity score; clustering, by the computer, each pairwise match identified in the plurality of candidate segments and the plurality of query segments, wherein the computer updates the voice detection model based upon the clustering.

7 . The method according to claim 1 , further comprising:
identifying, by the computer, a pairwise match including the candidate segment of the first audio signal and the query segment of the second audio signal having a similarity score; and storing, by the computer, into a database, matched data comprising each pairwise match identified in the first audio signal and the second audio signal.

8 . The method according to claim 1 , wherein extracting the plurality of features for each segment of the plurality of segments includes extracting, by the computer, a set of posterior probabilities for each of the one or more features of extracted from the plurality of audio segments.

9 . The method according to claim 8 , further comprising:
extracting, by the computer, the plurality of features for each segment of the plurality of segments for a new audio signal; and extracting, by the computer, the set of posterior probabilities for each of the one or more features of extracted from the plurality of audio segments of the new audio signal by applying the voice detection model configured to nullify the probability of at least a portion of the new audio signal having the discovered keyword.

10 . The method according to claim 1 , further comprising obtaining, by the computer, the plurality of audio signals from a database configured to store the plurality audio signals,
wherein the computer determines, for each candidate segment of each audio signal stored in the database, the plurality of similarity scores corresponding to the plurality of query segments of the second audio signal.

11 . A system comprising:
a database comprising a non-transitory memory configured to store one or more voice detection models; and a computer in communication with the database and comprising a processor configured to:
generate a plurality of segments of a plurality of audio signals, including one or more candidate segments of a first audio signal and one or more query segments of a second audio signal;
extract a plurality of features for each segment of the plurality of segments;
determine for a candidate segment of the first audio signal a plurality of similarity scores corresponding to the plurality of query segments of the second audio signal, each similarity score indicating a similarity of the features of the candidate segment with respect to the features of a corresponding query segment of the second audio signal;
identify a discovered keyword included in the candidate segment, in response to determining that at least one similarity score of the candidate segment satisfies a pairwise matching threshold; and
update a voice detection model associated with the plurality of audio signals to nullify a probability of a portion of an audio signal having the discovered keyword.

12 . The system according to claim 11 , wherein when identifying the discovered keyword in the candidate segment the computer is further configured to determine that the similarity score for the candidate segment satisfies the pairwise matching threshold.

13 . The method according to claim 1 , wherein when generating the plurality of segments of the plurality of audio signals the computer is further configured to:
parse the candidate segment of the first audio signal into a plurality of candidate paths, each candidate path comprising a fixed portion of the candidate segment; and parse the query segment of the second audio signal into a plurality of test paths, each test path comprising the fixed portion of the query segment.

14 . The system according to claim 13 , wherein the computer is configured to determine the similarity score for the candidate segment based upon comparing the plurality of candidate paths of the candidate segment and the plurality of test paths in the query segment, and
wherein the computer is configured to identify the discovered keyword in the candidate segment based upon determining that the similarity score for the candidate segment satisfies the pairwise-match threshold score.

15 . The system according to claim 14 , wherein when identifying the discovered keyword in the candidate segment the computer is further configured to determine that a threshold number of candidate paths of the candidate segment satisfy a second threshold score.

16 . The system according to claim 11 , wherein the computer is further configured to:
identify a pairwise match including the candidate segment of the first audio signal and the query segment of the second audio signal having a similarity score; cluster each pairwise

Assignees

• Pindrop Security Inc

Inventors

• Rao Hrishikesh

Classifications

• G10L15/197 Primary

  Probabilistic grammars, e.g. word n-grams · CPC title

• G10L2015/088

  Word spotting · CPC title

• G10L15/30

  Distributed recognition, e.g. in client-server systems, for mobile phones or network applications · CPC title

• G10L15/22

  Procedures used during a speech recognition process, e.g. man-machine dialogue · CPC title

• G10L25/54

  for retrieval · CPC title