US 12,236,438 B2
System and method for real-time fraud detection in voice biometric systems using phonemes in fraudster voice prints
Matan Keret, Oulu (FI); Roman Frenkel, Ashdod (IL); and Zvika Horev, Ra'anana (IL)
Assigned to Nice Ltd., Ra'anana (IL)
Filed by NICE Ltd., Ra'anana (IL)
Filed on Jan. 4, 2022, as Appl. No. 17/568,408.
Prior Publication US 2023/0214850 A1, Jul. 6, 2023
Int. Cl. G06Q 30/018 (2023.01); G06F 16/65 (2019.01); G10L 15/02 (2006.01); G10L 17/06 (2013.01)
CPC G06Q 30/0185 (2013.01) [G06F 16/65 (2019.01); G10L 15/02 (2013.01); G10L 17/06 (2013.01); G10L 2015/025 (2013.01)] 14 Claims
OG exemplary drawing
 
1. A computer-implemented method for proactive fraudster exposure in a customer service center having multiple service channels, the computer-implemented method comprising:
receiving incoming audio call interactions;
storing the audio call interactions in an interactions database;
generating a voiceprint for each of the incoming audio call interactions;
grouping the audio call interactions in the interactions database into one or more clusters of voiceprints based on the voiceprints generated for the interactions, wherein each one of the one or more clusters is associated with a speaker's voice based on the voiceprints;
for each pair of voiceprints in a voiceprint cluster:
extracting a first phoneme sequence for a first voiceprint of the pair of voiceprints and extracting a second phoneme sequence for a second voiceprint of the pair of voiceprints;
calculating a similarity score between the extracted first phoneme and the extracted second phoneme representation;
determining if a match exists between the extracted first phoneme sequence and the extracted second phoneme sequence based on a threshold; and
if a match exists, adding the matched phoneme sequence to a watchlist;
wherein the similarity score is calculated by:
creating a first vector for the extracted first phoneme sequence and a second vector for the extracted second phoneme sequence; and
comparing, between the first vector and the second vector in an indexed i×j matrix, a Levenshtein distance for each i×j element of the matrix between a phoneme sequence of length i of the first vector and a phoneme sequence of length j of the second vector, wherein the distance is a cost to modify the first subset to the second subset, the comparing comprising:
for each i×j element of the matrix, compare the distance values for three locations, (i, j−1), (i−1, j), and (i−1, j−1);
if the terminating phoneme of the first subset at index i is equivalent to the terminating phoneme of the second subset at position j, then the i×j element equals the minimum value of the three locations;
if the terminating phoneme of the first subset at index i is not equivalent to the terminating phoneme of the second subset at position j, then the i×j element is the minimum value of the three locations plus a cost; and
wherein the similarity score is a final row and final column element of the matrix.