| Summary: | 碩士 === 國立成功大學 === 電機工程學系 === 102 === This paper presents a novel coherence-function-based noise suppression algorithm (NSA) with a weighted overlap-add (WOLA) filterbank for dual microphones. It solved the following two issues: One is that traditional method cannot efficiently suppress the noise from the front, and the other is that it may cause estimation errors while the noise is from the back. Consider the complexity in algorithm; the proposed method employs a simple sound source localization (SSL) algorithm, and an awareness computation (AC) strategy to improve these drawbacks instead of using complex voice active detector (VAD). By calculating the cross-correlation of dual microphones, the information of time difference of arrival (TDOA) is obtained. Hence, the direction of noise source can be effectively estimated. To reduce the error rate of finding out the exact noise source, an adaptive threshold value is introduced. From the view of system integration, the AC and module-sharing scheme are also adopted to reduce the computational complexity. The results show that the number of multiplication of the proposed method is only 9.09% increased, and the SNR of the proposed algorithm has at least 3dB growth which is higher than that of other approaches. In FPGA implementation, the proposed SSL design can be operated at 25 MHz which is easily to achieve the real-time requirement of 72.625 kHz. Overall, it is very suitable for integrating with Fourier-transform-based WOLA hearing aid design in the future.
|
|---|
