Figure 2.

Comparison of several linear methods for direct decoding of acoustic features of speech. (A) Pearson correlations of decoded F0 from linear methods. Ridge and linear regressions were trained on the first 100 PCA components of neural activity and PLS regression was trained with 12 components. Ridge regressions were trained using 3 different methods to compute the λ factor: L-curve (L), cross-validation (X), and cross-validation with individual λ per features (Xm). (B) Pearson correlations of decoded F0 using a linear regression trained on varying PCA reductions of neural activity. Statistical significance computed by Quade-Conover test [Quade test: p < 0.001, t_(4,2516) = 834.8]. (C) Pearson correlations of decoded Mel cepstrum from linear methods. Ridge and linear regressions were trained on the first 100 PCA components of neural activity and PLS regression was trained with 12 components. Ridge regressions were trained using 3 different methods to compute the λ factor: L-curve (L), cross-validation (X) and cross-validation with individual λ per features (Xm). (D) Pearson correlations of decoded Mel cepstrum using a linear regression trained on varying PCA reductions of neural activity. Statistical significance computed by Quade-Conover test [Quade test: p < 0.001, t_(4,2516) = 837.4]. Conover comparisons significance for (B, D): n.s: p ≥ 0.05 [(D): t₍₂₅₁₆₎ = 0.5], *p = 0.013 [(B): t₍₂₅₁₆₎ = 2.5], ***p < 0.001 [(B, D): t₍₂₅₁₆₎ > 6.5].