Fig. 5
A comparison of average PESQ and STOI scores of TIMIT speech signals in practical scenarios with iSNR=0. The vertical margins exhibit the confidence (standard deviation) intervals computed over 24 speech utterances. hmax,prac(k,n) and \(\widetilde {\mathbf {h}}_{\mathrm {max,prac}}(k,n)\) denote, respectively, the linear and quadratic maximum SNR filters in practical conditions; hmax(k,n) and \(\widetilde {\mathbf {h}}_{\text {max}}(k,n)\) denote, respectively, the linear and quadratic maximum SNR filters in ideal conditions; MMSE-LSA denotes the log-spectral amplitude estimator of [3]; STSS denotes the spectral subtraction in the short-time modulation domain method of [30]; Y(k,n) denotes the input noisy observation: a PESQ scores and b STOI scores. We set N=3 for all four maximum SNR filters and Ď„y=2 for the two quadratic maximum SNR filters