US 11,056,118 B2
Speaker identification
David Martínez González, Saragossa (ES); and Carlos Vaquero Avilés-Casco, Madrid (ES)
Assigned to Cirrus Logic, Inc., Austin, TX (US)
Filed by Cirrus Logic International Semiconductor Ltd., Edinburgh (GB)
Filed on Jun. 28, 2018, as Appl. No. 16/21,637.
Claims priority of provisional application 62/526,558, filed on Jun. 29, 2017.
Prior Publication US 2019/0005962 A1, Jan. 3, 2019
Int. Cl. G10L 17/14 (2013.01); G10L 17/02 (2013.01); G10L 17/12 (2013.01)
CPC G10L 17/14 (2013.01) [G10L 17/02 (2013.01); G10L 17/12 (2013.01)] 17 Claims
OG exemplary drawing
 
1. A method of speaker identification, comprising:
receiving a speech signal;
dividing the speech signal into segments; and
following each segment:
extracting a plurality of features from a most recently received segment;
deriving scoring information comprising sufficient statistics of the extracted features of the most recently received segment;
combining the scoring information derived from the extracted features of the most recently received segment with previously stored scoring information comprising sufficient statistics of the extracted features of any previously received segment to form updated combined scoring information comprising sufficient statistics of the entire received speech signal to date, wherein the updated combined scoring information requires the same number of bits of memory as the previously stored scoring information;
storing the updated combined scoring information; and
calculating an identification score by comparing the updated combined scoring information with a previously stored model of speech of an enrolled speaker, such that the calculated identification score indicates a likelihood that the received speech signal represents the speech of the enrolled speaker.