公开(公告)号：US12198714B2

公开(公告)日：2025-01-14

申请号：US17850906

申请日：2022-06-27

Applicant: Industrial Technology Research Institute

Inventor： Liang-Hsuan Tai ,  Hong-Yu Chen ,  Yen-Ting Wu ,  Ting-Yu Wang

IPC: G10L25/30 ,  G10L17/26 ,  G10L21/0208 ,  H04B1/40

Abstract: The disclosure relates to a voice signal analysis method and device and a chip design method and device. The voice signal analysis method includes: in a first updating gradient, training a resolution recovery model by using first voice training data meeting a same grouping condition in multiple mission sets; in a second updating gradient, training the resolution recovery model by interleavingly using second voice training data meeting different grouping conditions in the mission sets; iteratively executing the first and second updating gradients to set an initial model parameter of the resolution recovery model; and recovering a high-resolution snore signal from a low-resolution snore signal by using the resolution recovery model. The low-resolution snore signal has a lower resolution than the high-resolution snore signal.

公开(公告)号：US20230368810A1

公开(公告)日：2023-11-16

申请号：US17850906

申请日：2022-06-27

Applicant: Industrial Technology Research Institute

Inventor： Liang-Hsuan Tai ,  Hong-Yu Chen ,  Yen-Ting Wu ,  Ting-Yu Wang

IPC: G10L25/30 ,  G10L21/0208 ,  G10L17/26 ,  H04B1/40

CPC classification number: G10L25/30 ,  G10L21/0208 ,  G10L17/26 ,  H04B1/40

Abstract: The disclosure relates to a voice signal analysis method and device and a chip design method and device. The voice signal analysis method includes: in a first updating gradient, training a resolution recovery model by using first voice training data meeting a same grouping condition in multiple mission sets; in a second updating gradient, training the resolution recovery model by interleavingly using second voice training data meeting different grouping conditions in the mission sets; iteratively executing the first and second updating gradients to set an initial model parameter of the resolution recovery model; and recovering a high-resolution snore signal from a low-resolution snore signal by using the resolution recovery model. The low-resolution snore signal has a lower resolution than the high-resolution snore signal.