| Summary: | 碩士 === 國立中正大學 === 電機工程所 === 95 === Abstract
Speaker identification is a kind of biological authentication technology, which uses personal characteristics to distinguish users. This technology becomes more and more important as the recognition technology matures. Since only a small number of voice features has been used to characterizes human voice, and the performance of speaker identification is easily disturbed by noise and personal health and emotions. Traditional speaker identification must apply recognition technologies in noiseless environments for high performance. This thesis proposes a novel glottal vibration feature to improve the traditional speaker identification system.
This study stems from the fact that human sound is caused by the vibration of vocal cord. Therefore, we extract both the features of the sound and glottal motion and compares their differences. The glottal motion was proven to be a good characteristic of human based on the features and the proposed classifier in the experiments. The results show that high accuracy of 92% was obtained with only glottal motion signal. The accuracy is increased to 96% with combined traditional speaker identification and glottal motion. Even in noisy environments, using glottal signal can achieve an accuracy of 88%. In the final experiment, we also study the effect of noise to the glottal-signal-assisted speaker identification system in different signal to noise ratio. The propose system is demonstrated to be effecting and noise resistant.
|
|---|
