On DCT-based MMSE estimation of short time spectral amplitude for single-channel speech enhancement

Log-Spectral Amplitude and Spectral Polarity Estimation in Short-Time Discrete Cosine Transform Domain

“…This means when the ISNR is above 0 dB, leaving the DCT polarity unmodified has no effect on perceived speech quality; however, an accurate DFT phase estimation might be required to achieve the same improvement in perceived speech quality. This advantage was further proved through subjective listening tests and the average of the scores given by the listeners, termed as mean subjective preference (%) score, was used as an indicator for the perceived speech quality [19]. Explicitly, for all tested noise conditions, using noisy PoS for speech reconstruction achieved a higher preference score than using the noisy PhS (Fig.…”

“…It is important to note that, DFT-based estimators cannot simply be carried over to the DCT domain as DCT is a real-valued transform and DFT is complex. In [19]. The polarity-only (PO) [or phase-only (PhO)] stimuli was generated by adding a controlled level of distortion into the DCT polarity spectrum (or the DFT phase spectrum), while keeping its spectral amplitudes fixed from the clean input.…”

“…this regard, DCT-based MMSE spectral amplitude (STSA) estimators were developed in [19] by minimizing the MSE between the original STSA and its estimator. To be used in conjunction with the DCT-based STSA estimators, an optimal MMSE estimator of the DCT noise power was also derived.…”

“…While the distortion measure previously used in [19] leads to good results, it is not the most perceptually meaningful one. Studies show that the human auditory performs a logarithmic compression of the STSA [20], and thus the distortion measure which is based on the MSE of log-spectral amplitude (LSA) is more perceptually relevant than that of the STSA.…”

“…Accordingly, the effect of using PoE on perceived speech quality is examined. The performance of our new estimators is compared to the DCT STSA estimator derived in [19], some of the well known DFT-based LSA estimators, e.g., [4], [6], as well as the State-Of-The-Art (SOTA) DFT-based phase-aware system, e.g., [25].…”