US 12,477,273 B2
Beamforming method and beamforming system using neural network
Kang Hun Ahn, Daejeon (KR); and Sang-Hyun Park, Daegu (KR)
Assigned to Deep Hearing Corp., Daejeon (KR)
Appl. No. 18/035,297
Filed by DEEPHEARING INC., Daejeon (KR)
PCT Filed Sep. 29, 2021, PCT No. PCT/KR2021/013328
§ 371(c)(1), (2) Date May 4, 2023,
PCT Pub. No. WO2022/097919, PCT Pub. Date May 12, 2022.
Claims priority of application No. 10-2020-0146191 (KR), filed on Nov. 4, 2020.
Prior Publication US 2023/0269532 A1, Aug. 24, 2023
Int. Cl. H04R 3/00 (2006.01); G06F 17/14 (2006.01); G06N 3/0464 (2023.01); G06N 3/08 (2023.01); G06N 3/084 (2023.01); H04B 7/06 (2006.01); H04N 7/15 (2006.01); H04R 1/40 (2006.01)
CPC H04R 3/005 (2013.01) [G06F 17/14 (2013.01); G06N 3/08 (2013.01); H04R 1/406 (2013.01); G06N 3/0464 (2023.01); G06N 3/084 (2013.01); H04B 7/0617 (2013.01); H04N 7/15 (2013.01); H04R 2201/401 (2013.01); H04R 2430/03 (2013.01); H04R 2430/23 (2013.01)] 6 Claims
OG exemplary drawing
 
1. A beamforming method, comprising the steps of:
receiving, respectively, a first sound signal and a second sound signal using a first microphone and a second microphone disposed apart from the first microphone by a predetermined distance;
obtaining Fourier transform results for the first sound signal and the second sound signal, respectively;
acquiring a phase difference between the first sound signal and the second sound signal from the Fourier transform results;
inputting only the phase difference into a beamforming model and performing neural network operations using a neural processor, wherein the neural processor performs the beamforming method independently or together with a processor that performs overall control of a beamforming device, and wherein the beamforming model is trained to perform the neural network operations based only on the phase difference;
performing element-wise multiplication between corresponding components of a matrix representing an output of the neural network operations and a matrix representing the Fourier transform results for the first sound signal; and
outputting the element-wise multiplication results,
wherein the neural processor obtains a mask using a Soft Binary Mask (SBM) method through the neural network operations.