METHOD AND SYSTEM FOR ENHANCING SPEECH BASED ON REINFORCEMENT LEARNING

Presented are a method for enhancing voice based on reinforcement learning and a system thereof. According to one embodiment of the present invention, the method may comprise: an initialization step of extracting a time-frequency gain value from a training data set and performing prerequisite learni...

Full description

Saved in:
Bibliographic Details
Main Authors JOON HYUK CHANG, TAE JUN PARK
Format Patent
LanguageEnglish
Korean
Published 02.09.2019
Subjects
Online AccessGet full text

Cover

Loading…
More Information
Summary:Presented are a method for enhancing voice based on reinforcement learning and a system thereof. According to one embodiment of the present invention, the method may comprise: an initialization step of extracting a time-frequency gain value from a training data set and performing prerequisite learning an action-value function; and a training step of obtaining an optimal time-frequency gain value selection policy by performing reinforcement learning using the action-value function learned through reward and improving a voice signal using the time-frequency gain value according to the optimal time-frequency gain value selection policy. 강화학습 기반의 음성 향상 방법 및 시스템이 제시된다. 일 실시예에 따른 강화학습 기반의 음성 향상 방법은, 훈련 데이터셋으로부터 시간-주파수 이득값을 추출하고, 행동 가치(action-value) 함수를 선행 학습시키는 초기화 단계; 및 보상(reward)을 통해 학습한 상기 행동 가치 함수를 이용하여 강화학습하여 최적의 시간-주파수 이득값 선택 정책을 획득하고, 상기 최적의 시간-주파수 이득값 선택 정책에 따른 시간-주파수 이득값을 이용하여 음성신호를 향상시키는 훈련 단계를 포함하여 이루어질 수 있다.
Bibliography:Application Number: KR20180051602