US 11,756,554 B2
Attribute identification method, and program
Hosana Kamiyama, Yokosuka (JP); Satoshi Kobashikawa, Yokosuka (JP); and Atsushi Ando, Yokosuka (JP)
Assigned to Nippon Telegraph and Telephone Corporation, Chiyoda-ku (JP)
Filed by Nippon Telegraph and Telephone Corporation, Chiyoda-ku (JP)
Filed on Aug. 23, 2021, as Appl. No. 17/408,597.
Application 17/408,597 is a continuation of application No. 16/606,986, granted, now 11,133,012, previously published as PCT/JP2018/018388, filed on May 11, 2018.
Claims priority of application No. 2017-103384 (JP), filed on May 25, 2017.
Prior Publication US 2021/0383812 A1, Dec. 9, 2021
This patent is subject to a terminal disclaimer.
Int. Cl. G10L 17/04 (2013.01); G06F 17/18 (2006.01); G10L 15/08 (2006.01); G06N 20/00 (2019.01)
CPC G10L 17/04 (2013.01) [G06F 17/18 (2013.01); G10L 15/08 (2013.01); G06N 20/00 (2019.01)] 4 Claims
OG exemplary drawing
 
1. An attribute identification method in an attribute identification device, wherein
a set of classes for identifying a speaker of uttered speech is to be an attribute,
the attribute identification method comprises:
calculating, from input speech, a posteriori probability sequence which is a sequence of posteriori probabilities that each frame of the input speech is a class c,
calculating, from the posteriori probability sequence, a value r(c) of the class c, the value r(c) indicating an extent to which the class c is a correct attribute identification result,
obtaining, from the value r(c), a most probable estimated class c  which is a class that is estimated to be a most probable attribute, and
deciding, based on the value r(c ) of the most probable estimated class c , whether the most probable estimated class c  is an attribute identification result or the most probable estimated class c  is to be rejected.