Wyniki 1-1 spośród 1 dla zapytania: authorDesc:"Nurbapa MEKEBAYEV"

Speech Recognizer-Based Non-Uniform Spectral Compression for Robust MFCC Feature Extraction DOI:10.15199/48.2018.06.17

Czytaj za darmo! »

The performance of speech recognition degrades dramatically in the presence of noise, due to spectral mismatch between the training and testing data. Therefore, robust speech recognition in noisy environment is still a challenging problem. To solve this problem, many compensation techniques have been proposed by researchers. In general, a compensation technique can be applied in the signal, feature or model space [1]. This paper focuses on the compensation in feature domain. Spectral compression is an effective robust feature extraction technique to reduce the mismatch between training and testing data in feature domain. In conventional Mel-frequency cepstral coefficients (MFCC) feature extraction, a logarithm function is applied to Mel filter bank energies in order to reduce their dynamic range. Root cepstral analysis [2] replace log function with a constant root function and yields RCC coefficients. RCC coefficients have shown better robustness against the noise. In RCC method compressed speech spectrum is computed as shown in (1): (1) PC (m)  P(m) , 0    1 where PC (m) is the compressed spectrum, P(m) is the original spectrum,  is the compression factor and m is the filter bank index. In (1), the compression factor is fixed for all the frequency bands under the assumption that the noise contamination is same throughout all frequency bands, although real world noise is mostly colored and does not affect the speech signal uniformly over the entire spectrum. Therefore, the compression factor should be adjusted for each band. Also, from the psychoacoustic point of view, using constant compression root for all frequencies is suboptimal [3]. Therefore, relation (1) is extended as follows: (2) P (m)  P(m) (m) , 0  (m)  1 C   where the compression factor is dependent on the frequency band and named non-uniform spectral compression. Even i[...]

 Strona 1