Nonintrusive Quality Assessment of Noise Suppressed Speech With Mel-Filtered Energies and Support Vector Regression

Effectiveness of ideal ratio mask for non-intrusive quality assessment of noise suppressed speech

“…Hence, some acoustic features must be used in order to capture the perceptual content of the speech signal under consideration. In this context, 40-D Mel Filterbank Energies (FBEs) are used as acoustic features [8]. Initially, we extract the true IRM features to show their effectiveness for quality assessment task using a T-F representation of clean speech signal.…”

“…The subjective scores of training utterances are used while training the regression model. We have used the mean of FBEs, which is proven to perform better than using the variance of FBEs [8] and mean and variance of the IRM features at a time. However, in non-intrusive quality assessment, the clean speech signal is not available.…”

“…Spectral dynamics, spectral flatness, spectral centroid, spectral variance, fundamental frequency or pitch (F 0 ) excitation variance and perceptual linear prediction (PLP) coefficients were used for quality prediction in [5], [6]. In [7] and [8], the quality assessment problem is posed as a regression problem and the mapping between acoustic features and the subjective score was found using Mel Frequency Cepstral Coefficients (MFCCs) and filterbank energies, respectively. To find the mapping, Support Vector Regression (SVR) was used.…”

“…Since IRM contains the information regarding the relative presence of residual noise in enhanced or noise suppressed speech, it provides a useful clue about the quality of a given utterance. The features extracted from IRM are used in addition to the standard acoustic feature set, namely, Mel Filterbank Energies (FBEs) to train a regression model and thus, the problem is posed as a regression problem [7], [8], [16]. The statistics of FBEs and IRM features are used as an input to the regression model.…”

Design of a Low Power and Area Efficient Architecture for the Detection of Audio Biological Signals

“…Furthermore, a floating-point MAC unit and memories are shared with many processes to reduce hardware complexity and energy consumption remarkably but at the cost of operating Speed. [3] and [4] have presented a new method for non-intrusive quality assessment of noise-suppressed speech, by using mel-filter bank energies as features to capture signal variations, and SVR for feature mapping. We showed that noise injection and suppression affects the FBEs and such changes (represented by the mean and variances) are also effective and parameterizable to assess quality.…”