Dynamic Encoding of Acoustic Features in Neural Responses to Continuous Speech
Dynamic Encoding of Acoustic Features in Neural Responses to Continuous Speech
Article Figures & Data
Figures
-
- Download figure
- Open in new tab
- Download powerpoint
Figure 1.
Representation of phonetic information in the PRP. A , PRPs are calculated by averaging the time-aligned neural responses to all instances of a phoneme. B , The average EEG response to all phonemes averaged over all electrodes. Individual subjects are shown in gray; grand average PRP is shown in red. Time points where PRP shows a significant response are shaded in yellow (central electrode Cz, t test, multiple-comparison corrected via FDR, q < 0.01). The average acoustic spectrogram of all phonemes is shown on the left side. The scalp topographies of three significant time points based on Figure 1 B including 70, 130, and 200 ms are shown at the bottom. C , Grand average PRPs for 30 American English phonemes. PRPs of a frontocentral electrode FCz are plotted from −100 ms before phoneme onset to 600 ms after phoneme onset. D , Hierarchical clustering of PRPs using all electrodes shows encoding of phonetic information largely driven by manner of articulation, highlighted by different colors.
-
- Download figure
- Open in new tab
- Download powerpoint
Figure 2.
Time course of phonetic feature representation in the PRP. A , The average responses of phonemes that share the same manner of articulation show the time course of manner-specific PRPs (electrode FCz). B , F statistic for manner of articulation groups reveals three distinct intervals with significantly separable responses to different manners of articulation (shown in blue, FDR corrected, q < 0.05). Neural F statistic (black; response) is based on PRP responses recorded from electrode FCz. Acoustic F statistic (purple; stimulus) is based on acoustic spectrograms of phonemes. C , Scalp topographies of the F statistic calculated by each electrode for the four response events, R1, R2, R3, and R4. The two temporal electrodes of T7 and T8 are marked on the first topographical map. D , Similarity of confusion patterns for manners of articulation for acoustic and neural signals (r = 0.59, p = 0.016). E , Effect size accumulated over subjects and stimulus duration (F-statistic measure for electrode FCz).
-
- Download figure
- Open in new tab
- Download powerpoint
Figure 3.
Patterns of phoneme similarity in EEG. A , PRP similarity matrices at time points corresponding to R1, R2, R3, and R4. Similarity is defined as m − (Euclidean distance) where m is the mean value of each matrix. B , One-dimensional representation of the PRP distance matrices at each time point based on electrodes showing significant distinctions between manners of articulation (multiple-comparison corrected via FDR, q < 0.001). Time points where no significant electrode was found are set to zero.
-
- Download figure
- Open in new tab
- Download powerpoint
Figure 4.
Recurrent appearance of patterns of phoneme similarity in PRPs. A , Neural–acoustic similarity matrix: the covariance between acoustic similarity matrices and PRP similarity matrices at different time lags. The four distinct temporal events are marked as R1, R2, R3, and R4. B , Neural–neural similarity matrix: the covariance of PRP similarity matrices at different time lags.
-
- Download figure
- Open in new tab
- Download powerpoint
Figure 5.
Encoding of speaker characteristics in PRPs. A , Correlation between speaker-dependent pairwise phoneme distances in neural space and acoustic space. Maximum correlation occurs ∼200 ms after phone onset, corresponding to R3. The black contour indicates significant correlation (FDR corrected, q < 0.05). B , Scalp topography showing average speaker-dependent correlation from 190 to 230 ms. C , Two-dimensional MDS diagram of PRPs. Colored bubbles show manner of articulation; black letters denote speaker 1; white letters denote speaker 2.
-
- Download figure
- Open in new tab
- Download powerpoint
Figure 6.
Joint neural encoding of speaker and phonetic features. A , Significant correlation between speaker feature information encoded in PRPs and acoustic spectrograms. B , Average PRPs corresponding to vowels articulated by speaker 1 versus vowels articulated by speaker 2. SE is shown by shaded color based on different subjects. C , Three-dimensional MDS diagram showing the separation of PRPs between the two speakers. Silhouette index quantifies the clustering of PRPs using the following: S1, separation of manners of articulation in all PRPs; S2, separation of speakers in all PRPs; and S3, separation of speakers within group of vowels. D , Comparison of timing and the location of response for average PRP, response components correlated with acoustic discriminability of phoneme categories, and response components correlated with speaker differences. Electrodes proceed in the anterior direction in rows from left to right and are ordered as follows: frontal (F), central (C), parietal (P), occipital (O). Temporal electrodes are marked with T7 and T8.
valentinesaftessithe.blogspot.com
Source: https://www.jneurosci.org/content/37/8/2176/tab-figures-data
0 Response to "Dynamic Encoding of Acoustic Features in Neural Responses to Continuous Speech"
Post a Comment