A Study of Text Dependent Speaker Verification
碩士 === 國立中興大學 === 電機工程學系 === 91 === Dynamic time warping (DTW) algorithm was widely used in speech recognition but it take large computation time and difficult to determine the thresholding value. Hidden Markov Models (HMM) provides a natural and highly reliable way of recognizing speech for a wide...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Others |
| Language: | zh-TW |
| Published: |
2003
|
| Online Access: | http://ndltd.ncl.edu.tw/handle/32485095758346365459 |
| Summary: | 碩士 === 國立中興大學 === 電機工程學系 === 91 === Dynamic time warping (DTW) algorithm was widely used in speech recognition but it take large computation time and difficult to determine the thresholding value. Hidden Markov Models (HMM) provides a natural and highly reliable way of recognizing speech for a wide range of applications but it is too complex and too time consuming. In the thesis, we take some characteristics from DTW and HMM, and using the Gaussian distribution as front-end processes which can provide a good performance for voice verification.
In speaker verification, we take a voice from an unknown speaker to match a set of known speakers from database. Feature vectors are extracted from the voice samples by using Linear predictive coding (LPC) algorithm or Mel-Frequency Cepstral Coefficients (MFCC) algorithm. We compare the outcome of different feature vectors in both LPC and MFCC algorithm and to observe the influence of the state size. The simulation results show that using the LPC algorithm is better than using the MFCC algorithm in terms of the correctness to identify the right person with right password.
|
|---|