원문정보
초록
영어
Speaker Recognition and Verification is becoming one of the widely used forms of biometric authentication in today’s scenario where remembering strings of textual passwords and numbers are becoming a hassle. Authentication of users using voice offers many advantages and easy to use techniques. In this paper a comparison is drawn among the most commonly used feature extraction techniques in Speaker Recognition and Verification. Extracting useful and unique features from the user’s voice forms the backbone of an efficient Speaker Recognition System. Here, the most commonly used methods for Feature Extraction viz. MFCC (Mel Frequency Cepstral Coefficient), LPC (Linear Predictive Coefficient), PLP (Perceptual Linear Prediction) are discussed, compared and an attempt is made to deduce which one performs best.
목차
1. Introduction
2. Steps in Speaker Recognition
3. Feature Extraction
4. MFCC (Mel Frequency Cepstral Coefficient)
4.1. Framing
4.2. Fast Fourier Transform
4.3. Mel Scale Filtering
4.4. Logarithm
4.5. Discrete Cosine Transform
4.6. Delta Energy and Spectrum
5. LPC (Linear Predictive Coding)
5.1. Preemphasis
5.2. Frame Blocking
5.3. Windowing
5.4. Autocorrelation Analysis
5.5. LPC Analysis
5.6. Conversion of LPC Parameters to Cepstral Coefficients
6. PLP (Perceptual Linear Prediction)
6.1. Windowing
6.2. Calculation of Power Spectrum
6.3. Application of Frequency Warping into Bark Scale
6.4. Equal Loudness Pre Emphasis
6.5. Intensity Loudness
6.6. Linear Prediction
6.7. Cepstrum Computation
7. Tabular Comparison of MFCC, LPC and PLP
8. Conclusion
References