| Summary: | 碩士 === 國立中山大學 === 電機工程學系研究所 === 106 === In recent years, voice recognition has been widely used in many smart home and smart phone applications. By recognizing human being’s voices, more convenient operations are enabled. Voice recognition also plays an important role in automotive electronics. Human beings’ can thus safely control navigation systems or air conditioners during their driving.
However, with the feature size shrinking of the semiconductor manufacturing technology, voice recognition circuits are more easily to be affected by external noises. This may result in unrecognition of voices, and therefore invalidates the system operations, and even raises safety concerns. As a result, it is very critical to accurately detect noises in voice signals, and accordingly eliminate them. This is also the research objective of this thesis.
In this thesis we employ the ideal and non-ideal additive Gaussian white noise (AWGN) to simulate various types of noises in voice signals, and accordingly analyze their impacts on the recognition rate. Compared with the developed methods in the literature, we propose a more accurate noise detection technique. Our experimental results show that the proposed technique can achieve more than 99% accuracy.
On the other hand, we also investigate the effectiveness of the previous methods in the literature for eliminating noises. We find that for ideal AWGN, the recognition rate can be enhanced by only 4.69%, while for non-ideal AWGN, 8.15% enhancement is achieved. However, by integrating the proposed noise detection technique with the previous method to develop a new noise elimination technique, we find that 12% enhancement on the recognition rate for ideal AWGN can be achieved, while the enhancement for the non-ideal AWGN is 42.11%.
One major problem for the previous noise elimination methods is that the computation is quite complicated where the required execution time is 7.4 times the length of the target voice signal. In order to reduce the required computation complexity for eliminating noises, we further propose a linear approximation based novel noise elimination technique. The execution time of the proposed technique is only 6.5% of that for the previous method. For ideal AWGN, the proposed technique can enhance the recognition rate by 12%. As for non-ideal AWGN, the enhancement is 53.31%.
|
|---|
