Figure 3.

Direct decoding of acoustic features of speech using PLS regression. (A) Pearson correlations of decoded Mel cepstral coefficients (blue) and the corresponding chance levels (red) using a PLS regression with 12 components, 210 ms of time context and 0 ms of time delay. (B) Average Pearson correlations of decoded Mel cepstrum across coefficients and Pearson correlations of decoded F0 using a PLS regression with 12 components, 210 ms of time context and 0 ms of time delay. (C) Pearson correlations of decoded Mel cepstrum with varying time delay using PLS regression with 12 components and 210 ms of time context. (D) Pearson correlations of decoded Mel cepstrum with varying time context using PLS regression with 12 components and 0 ms of time delay. (E) Pearson correlations of decoded Mel cepstrum using PLS regression with varying number of components, 210 ms of time context and 0 ms of time delay. Statistical significance with respect to chance levels computed with Bonferroni-corrected Wilcoxon signed rank test for (A, B) (see values in Section 3.2). Statistical significance computed by Quade-Conover test for (C) [Quade test: p < 0.001, t_(4,2516) = 399.8], (D) [Quade test: p < 0.001, t_(3,1887) = 1388.7] and (E) [Quade test: p < 0.001, t_(7,4403) = 254.9]. Conover comparisons for (C–E): n.s: p ≥ 0.05 [(E): t₍₄₄₀₃₎ < 1.7], ***p < 0.001 [(C): t₍₂₅₁₆₎ > 4.6, (D): t₍₁₈₈₇₎ > 18.8, (E): t₍₄₄₀₃₎>7.3]. Arrows indicate best accuracies.