| Summary: | 碩士 === 淡江大學 === 電機工程學系碩士班 === 104 === The speaker recognition is always a hot topic in the research field. Technologies of speaker recognition under white and color noisy environments have been proposed in recent years. Sparse representation algorithm has been introduced into noise filtering for improving the assessments of speech quality, such as SNR, SNRseg, LLR and PESQ, but the cost time is lengthy. So we employ Label Consistent K-SVD sparse coding (LC-KSVD) to de-noise speech data and decrease processing time. Speaker recognition systems almost use Euclidean distance to compute the distance between features, currently. Our goal is to have short corpus and independent corpus, which makes it more difficult to achieve high recognition accuracy. We propose Riemannian distance replace Euclidean distance, but our experimental results show that Euclidean distance is superior than Riemannian distance. We use waveform, MFCC and MFCC smoothing spectrum with RD and ED for speaker recognition experiment in this paper.
|
|---|
