TY  - JOUR
TT  - Speech enhancement with spectral subtraction based kalman filter
AU  - Kutlu, Cem
AU  - Arserim, Muhammet Ali
PY  - 2012
DA  - June
JF  - Dicle Üniversitesi Mühendislik Fakültesi Mühendislik Dergisi
JO  - DUJE
PB  - Dicle University
WT  - DergiPark
SN  - 1309-8640
SP  - 49
EP  - 59
VL  - 3
IS  - 1
KW  - Ses iyileştirme
KW  - Kalman filtresi
KW  - Spektral çıkarma
N2  - Speech enhancement techniques aim to improve the quality or intelligibility of speech signals contaminated with background noise and can be implemented both in time and frequency domains. Spectral subtraction, one of the most feasible methods in practice, is an effective technique to enhance the noisy speech signals. However, a residual noise called musical noise occurs with the estimated speech signal and this is the major inconvenience of Spectral subtraction. Wiener filter is an alternative approach for speech enhancement in the manner of Spectral subtraction filter. The drawback of the Wiener filter is the fixed frequency response at all frequencies and the requirement to estimate the power spectral density of the clean signal and the noise prior to filtering. Kalman filtering is also one of the most effective methods in speech enhancement. In recent years, due to its magnificent accurate estimation characteristics especially in the research field of navigation and GPS, researchers tried to manipulate its advantages for useful purposes in signal processing. However, to improve the speech signals with the Kalman filter, some parameters such as the autoregressive (AR) coefficients of the clean signal and the noise covariance matrix must be known. Determining the AR coefficients of clean speech signal plays a crucial role for the success of the Kalman filter while the only noisy observations are available. In such condition it is very difficult to estimate these parameters and today researches on this issue are ongoing. In this study, the parameters necessary to implement the Kalman filter is determined using Spectral subtraction. First of all, Spectral subtraction, Wiener filter and Kalman filter is analyzed respectively. Then all three methods mentioned above are carried out for speech signals corrupted with different types of noise. Finally, Kalman filter combined with Spectral Subtraction proposed in this study is applied to those signals and all results are compared based on output SNR values as an objective measurement for the enhancement performance. The results achieved in this study has shown that, if the AR cofficients of the original signal and noise variance is known Kalman filter is sufficient alone for the enhancement of noisy speech signals. In addition, musical noise which occurs with the methods based on the noise spectrum estimation, does not occur with the Kalman filtering. However, the assumption that these parameters are known is not workable in practice. Therefore, the AR coefficients are calculated by using the signal enhanced with Spectral subtraction and also noise variance is calculated by subtracting the enhanced signal from the noisy signal. In the last stage, Kalman filter was applied to the noisy signal using the parameters determined with Spectral subtraction. Considering the obtained results, combined Kalman filter provided a better SNR improvement compared to the Wiener filter and Spectral subtraction. Also combined Kalman filter suppressed the musical noise that occurred owing to Spectral subtraction. When the SNR values are taken into account, it is seen that the Kalman filter alone provided a better SNR improvement than combined Kalman filter.This is because of using the original signal while calculating the AR parameters for the Kalman filter alone.
UR  - https://dergipark.org.tr/en/pub/dumf/issue//312686
L1  - https://dergipark.org.tr/en/download/article-file/302708
ER  -