Title:
TARGET SOUND EMPHASIS DEVICE, NOISE ESTIMATION PARAMETER LEARNING DEVICE, METHOD FOR EMPHASIZING TARGET SOUND, METHOD FOR LEARNING NOISE ESTIMATION PARAMETER, AND PROGRAM
Document Type and Number:
WIPO Patent Application WO/2018/110008
Kind Code:
A1
Abstract:
The present invention provides a noise estimation parameter learning device capable of emphasizing a target sound using a spectral subtraction method while coordinating a plurality of microphones arranged at distant locations even in a large space that tends to cause problematic reverberation or time frame difference. The noise estimation parameter learning device, which is for learning a noise estimation parameter that is used for the estimation of the noise included in an observation signal acquired from the plurality of microphones, comprises: a modeling unit for modeling the probability distribution of the observation signal from a prescribed microphone, modeling the probability distribution of a time frame difference, and modeling the probability distribution of a transfer function gain; a likelihood function setting unit for setting a likelihood function associated with the time frame difference and a likelihood function associated with the transfer function gain on the basis of the modeled probability distribution; and a parameter update unit for repeatedly updating a variable for the two likelihood functions alternately and outputting the time frame difference and the transfer function gain after convergence as a noise estimation parameter.
Inventors:
KOIZUMI YUMA (JP)
SAITO SHOICHIRO (JP)
KOBAYASHI KAZUNORI (JP)
OHMURO HITOSHI (JP)
SAITO SHOICHIRO (JP)
KOBAYASHI KAZUNORI (JP)
OHMURO HITOSHI (JP)
Application Number:
PCT/JP2017/032866
Publication Date:
June 21, 2018
Filing Date:
September 12, 2017
Export Citation:
Assignee:
NIPPON TELEGRAPH & TELEPHONE (JP)
International Classes:
G10L21/0264; G10L21/0232
Domestic Patent References:
| WO2007100137A1 | 2007-09-07 |
Other References:
S. BOLL: "Suppression of acoustic noise in speech using spectral subtraction", IEEE TRANS. ASLP, 1979
T. HIGUCHIH. KAMEOKA: "Joint audio source separation and dereverberation based on multichannel factorial hidden Markov model", PROC MLSP 2014, 2014
HIDEKI ASOH: "ShinSo GakuShu, Deep Learning", November 2015, KINDAI KAGAKU SHA CO., LTD.
See also references of EP 3557576A4
T. HIGUCHIH. KAMEOKA: "Joint audio source separation and dereverberation based on multichannel factorial hidden Markov model", PROC MLSP 2014, 2014
HIDEKI ASOH: "ShinSo GakuShu, Deep Learning", November 2015, KINDAI KAGAKU SHA CO., LTD.
See also references of EP 3557576A4
Attorney, Agent or Firm:
NAKAO, Naoki et al. (JP)
Download PDF: