Fig. 5.

Speech envelope tracking: For the 27 subjects indicated by different colours, the above plots show the power ratios for the DSS approach and the generalized eigenvalues (λ′s) for the SI-GEVD approach (averaged across trials) for the 63 components. The drops in the curves are sharper for the SI-GEVD case showing that fewer components are necessary to ensure high SNR at the output in comparison to the DSS case.

• 2)
  Results: Correlations from the raw, SI-GEVD filtered and DSS-filtered data were compared using Wilcoxon’s signed-rank test with $\alpha =0.05$ (Fig. 4). The correlations from the SI-GEVD filtering data were found to be significantly higher than those of raw data ($p=0.0082,W=81$). As can be seen in Fig. 5, power ratio plots from DSS showed a gradual decrease of power ratio over components, indicating that the power of the stimulus following responses were spread over more than a few components, and hence choosing a small K would be insufficient to ensure a high SNR in the filtered data, while choosing a larger K will capture too much of the noise. In Fig. 4, it can indeed be seen that SI-GEVD filtering results in correlations significantly higher than with DSS filtering ($p=0.0001,W=342$ for ‘DSS-3reps’).