Redukce šumu v řeči pomocí pole mikrofonů a neuronových sítí
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Tato práce se zabývá způsoby redukce šumu pomocí pole mikrofonů v reálném čase. Jsou zde ukázány způsoby zpracování vícekanálových záznamů z evaluační kampaně CHiME-4 a CHiME-5. Seznámíme se s problematikou odhadu šumu pomocí pole mikrofonů. Je zde ukázána důležitost potřeby detektoru aktivity řeči VAD. Pro realizaci redukce šumu v reálném čase jsou zde navrženy dva systémy. Jeden využívá metody beamformingu za pomocí VAD a druhý pracuje na principu slepé separace signálu, přesněji pak extrakci nezávislého vektoru. Oba Systémy jsou realizovány pomocí programovacího jazyka C++ a knihovny PortAudio. První systém se v praxi ověřil jako velmi efektivní, zatím co druhý systém fungoval pouze omezeně. U~obou systémů proběhla optimalizace s důrazem na minimální možnou hodnotu zpoždění a vytížení procesoru.
This thesis focuses on noise reduction in real-time using an microphone array. A number of techniques is shown being used in the processing of multichannel recordings from the CHiME-4 and CHiME-5 challenge. The solution to the issue of noise estimation using an microphone array is presented. Also the importance of a good VAD is shown. For the realization of noise reduction in real-time we designed two systems. One is based on beamforming aided by VAD, the other one is based on blind signal separation, more precisely on independent vector extraction. Both systems are implemented using the programming language C++ and PortAudio library. The first system is shown as highly effective, whereas the other one has limited functionality. Both were optimized with the emphasis on minimal latency and processor usage.
This thesis focuses on noise reduction in real-time using an microphone array. A number of techniques is shown being used in the processing of multichannel recordings from the CHiME-4 and CHiME-5 challenge. The solution to the issue of noise estimation using an microphone array is presented. Also the importance of a good VAD is shown. For the realization of noise reduction in real-time we designed two systems. One is based on beamforming aided by VAD, the other one is based on blind signal separation, more precisely on independent vector extraction. Both systems are implemented using the programming language C++ and PortAudio library. The first system is shown as highly effective, whereas the other one has limited functionality. Both were optimized with the emphasis on minimal latency and processor usage.
Description
Subject(s)
Redukce šumu, CHiME-4, VAD, real-time aplikace, Noise reduction, CHiME-4, VAD, real-time application