Title:
SPEAKER DIARIZATION METHOD, SPEAKER DIARIZATION DEVICE, AND SPEAKER DIARIZATION PROGRAM
Document Type and Number:
WIPO Patent Application WO/2022/123742
Kind Code:
A1
Abstract:
A speaker vector extraction unit (15b) uses a time series of audio features for each frame of a most recent audio signal to extract speaker vectors representing a speaker feature for each frame. A learning unit (15d) uses a speaker vector and a speaker label representing an estimated speaker for the speaker vector to generate, through learning, an online EEND model (14a) that estimates a speaker label for the speaker vector for each frame.
Inventors:
ANDO ATSUSHI (JP)
MURATA YUMIKO (JP)
MORI TAKESHI (JP)
MURATA YUMIKO (JP)
MORI TAKESHI (JP)
Application Number:
PCT/JP2020/046117
Publication Date:
June 16, 2022
Filing Date:
December 10, 2020
Export Citation:
Assignee:
NIPPON TELEGRAPH & TELEPHONE (JP)
International Classes:
G10L17/00; G10L15/04; G10L17/02; G10L17/04; G10L17/18
Foreign References:
| JP2019527370A | 2019-09-26 |
Attorney, Agent or Firm:
SAKAI INTERNATIONAL PATENT OFFICE (JP)
Download PDF: