Television Echo Cancellation with Microphone Array and Frequency-Domain Deep Recurrent Neural Networks

• Main Authors: Wei-Jung Hung, 洪瑋嶸
• Other Authors: 廖元甫
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/6turwe

碩士 === 國立臺北科技大學 === 電子工程系碩士班（碩士在職專班） === 105 === The purpose of this research was to study TV program echo cancellation under the situation of smart TV. We expect the receiver can receive the clean speech command without being disturbed by continuous playing TV program. The survey was conducted by several subjects. First, our echo cancellation system was deal with MVDR that can reach the ability of strengthened speech signal from the source pointed to USER and echo noise inhibition. Then, we use frequency-domain RNN as Adaptive Filter to cancel TV echo. Final, the remaining echo will be filtered out again by Spectral Subtraction Post-processing. The survey was simulated under the situation of various types of TV shows, speech, SNR and sides of speech by the way of ERLE to judge which is better. The result shows the way we mention of which average ERLE is 11.9dB rather than NLMS’s 6.1dB. The principal conclusion was that if we made MVDR and frequency domain RNN, the efficiency will be much more better than others and the noise can be truly eliminated effectively.