The ideal goal of every speech enhancement algorithm is to estimate the clean speech from the noi... more The ideal goal of every speech enhancement algorithm is to estimate the clean speech from the noisy observations without loss of its quality and intelligibility. Traditional approaches such as spectral subtraction, Wiener ltering, etc. perform noise removal in stationary noise conditions. The most common approach in the speech denoising is to estimate the clean speech by minimizing the risk corresponding to cost functions such as squared error, or perceptually motivated distortion measures like Itakura-Saito (IS) distortion. Since direct minimization of these risks are not possible, estimators of risk such as Stein's unbiased risk estimator (SURE) are minimized to obtain an estimate of clean speech. Here, we try to improvise the estimators derived using SURE by considering the prior probability distributions of speech. We derived the risk estimators assuming that speech signal follows Gaussian and Laplacian distributions.
Probing the large scale structures in the universe at larger redshifts by studying the fluctuatio... more Probing the large scale structures in the universe at larger redshifts by studying the fluctuations in the red-shifted 21 cm emission from neutral hydrogen atom (HI) at earlier epochs, has drawn attention quite for a long time. Studies conducted in this area has shown that the measurement of visibility correlation function at different frequencies and baselines allow a direct relationship with the power spectrum of density fluctuations of the HI distribution thereby providing a method for studying large scale structures at large red-shifts. Recent works has also explored the possibilities of using visibility correlation measure of the red-shifted 21 cm emission as an estimator for the Epoch of Re-ionization (EoR) power. However, the direct correlation of visibilities as obtained from the correlator will have the foreground and noise components which are to be removed. This work addresses the problem of foreground source and noise removal from the visibility correlation, making use of the fast decaying property of the visibility correlation of HI emission. Also, we have proposed a scheme for extracting the signal iteratively so that the waiting time for obtaining large chunks of data to estimate the weak HI emission signal can be avoided.
The ideal goal of every speech enhancement algorithm is to estimate the clean speech from the noi... more The ideal goal of every speech enhancement algorithm is to estimate the clean speech from the noisy observations without loss of its quality and intelligibility. Traditional approaches such as spectral subtraction, Wiener ltering, etc. perform noise removal in stationary noise conditions. The most common approach in the speech denoising is to estimate the clean speech by minimizing the risk corresponding to cost functions such as squared error, or perceptually motivated distortion measures like Itakura-Saito (IS) distortion. Since direct minimization of these risks are not possible, estimators of risk such as Stein's unbiased risk estimator (SURE) are minimized to obtain an estimate of clean speech. Here, we try to improvise the estimators derived using SURE by considering the prior probability distributions of speech. We derived the risk estimators assuming that speech signal follows Gaussian and Laplacian distributions.
Probing the large scale structures in the universe at larger redshifts by studying the fluctuatio... more Probing the large scale structures in the universe at larger redshifts by studying the fluctuations in the red-shifted 21 cm emission from neutral hydrogen atom (HI) at earlier epochs, has drawn attention quite for a long time. Studies conducted in this area has shown that the measurement of visibility correlation function at different frequencies and baselines allow a direct relationship with the power spectrum of density fluctuations of the HI distribution thereby providing a method for studying large scale structures at large red-shifts. Recent works has also explored the possibilities of using visibility correlation measure of the red-shifted 21 cm emission as an estimator for the Epoch of Re-ionization (EoR) power. However, the direct correlation of visibilities as obtained from the correlator will have the foreground and noise components which are to be removed. This work addresses the problem of foreground source and noise removal from the visibility correlation, making use of the fast decaying property of the visibility correlation of HI emission. Also, we have proposed a scheme for extracting the signal iteratively so that the waiting time for obtaining large chunks of data to estimate the weak HI emission signal can be avoided.
Uploads
the clean speech by minimizing the risk corresponding to cost functions such as squared error, or perceptually motivated distortion measures like Itakura-Saito (IS) distortion. Since direct minimization of these risks are not possible, estimators of risk such as Stein's unbiased risk estimator (SURE) are minimized to obtain an estimate of clean speech.
Here, we try to improvise the estimators derived using SURE by considering the prior probability distributions of speech. We derived the risk estimators assuming that speech signal follows Gaussian and Laplacian distributions.
the clean speech by minimizing the risk corresponding to cost functions such as squared error, or perceptually motivated distortion measures like Itakura-Saito (IS) distortion. Since direct minimization of these risks are not possible, estimators of risk such as Stein's unbiased risk estimator (SURE) are minimized to obtain an estimate of clean speech.
Here, we try to improvise the estimators derived using SURE by considering the prior probability distributions of speech. We derived the risk estimators assuming that speech signal follows Gaussian and Laplacian distributions.