Article Figures & Data
Figures
- Figure 1.
Illustration of the steps used to extract speech and EEG features and to estimate the association between them. The speech signal is passed through a gammatone filter bank simulating cochlear processing, and the envelope at the output of each filter (i.e., the envelope of each speech band) is extracted as a speech feature. Similarly, different bands of the EEG and different sensor channels together form the different EEG features. For the lower-frequency bands (delta and theta), the EEG signals are used as is. For the alpha and beta bands, both the signals in those bands, and their envelopes are extracted as separate features. For the higher-frequency gamma bands, only the envelopes of the EEG signals in those bands are considered. These EEG features are then compared with the speech features using spectral coherence.
- Figure 2.
Illustration of the effect of attention on the average speech-EEG magnitude squared coherence spectra, for (A) the envelope of the 1014-Hz speech band, and the low-frequency portions (overlapping with the delta and theta bands) of EEG channel C3, and for (B) the envelope of the 3733-Hz speech band, and the envelope of the low-gamma band of EEG channel CP1. Note that the y-axis ranges differ between A and B. The shaded regions indicate values within 1 SEM. The delta-band and theta-band EEG responses (A), and the low-gamma-band EEG envelope fluctuations (B), selectively track features of the attended speech over the ignored speech.
- Figure 3.
Differential effects of attention on speech-EEG coherences in different EEG bands (A), different speech bands (B), and the full matrix of EEG bands versus speech bands (C). A, Differential (attended–ignored) coherence averaged across speech bands and EEG channels (shown as a z-score) for each of the EEG bands. Uncorrected p values obtained from the permutation test are displayed for the different EEG bands. When a Bonferroni-corrected p value threshold of 0.05/8 = 0.006 is applied to each band, we find that the delta and theta bands show significantly higher coherence with speech when it is attended compared to when it is ignored. In addition, we also find that the envelope of the low-gamma band shows greater coherence with attended versus ignored speech. B, Differential coherence averaged across all EEG bands and EEG channels (shown as a z-score) for each input speech band. The strongest attention effects appear to occur in the 0.5- to 3-kHz range, which contains spectro-temporal speech features (formants and formant transitions) that convey many vowel and certain consonant cues, and is also the range thought to be the most important for speech intelligibility. In panel C, the differential coherence averaged across EEG channels is shown as a z-score for each EEG band and speech band for completeness. While the 0.5- to 3-kHz speech frequency range shows hot spots in the delta, theta, and low-gamma EEG bands, the lower-frequency speech bands (e.g., 200 Hz) show a hot spot only in the theta range corresponding to the syllabic rate. This could be because the pitch conveyed by the resolved harmonics of the syllabic voicing may be an important cue based on which attention is directed. In all three panels, z-scores shown are averaged across speech stories and individual subjects, with error bars representing the SE.
- Figure 4.
Scalp maps showing the average coherence (shown as a z-score) at each of the different EEG electrodes in the attended, ignored, and differential conditions. To obtain the scalp maps, the speech-EEG coherence values were averaged across the delta, theta, and low-gamma EEG bands (i.e., the bands showing significant attention effects in Fig. 3A), and all speech bands, and expressed as a z-score. The intensity shown at each electrode is the mean of the z-score across speech stories and individual subjects. Note that the scalp maps are scaled to their respective minimum and maximum z-score values, so as to best show the spatial patterns. The spatial profiles are hard to distinguish between the attended and ignored maps; however, note that the coherences are larger in the attended condition than the ignored, on an absolute scale. The differential map shown in the right column quantifies these differences across the scalp. Temporal-parietal regions appear to show the largest coherence differences between the attended and ignored conditions; however, this pattern is not symmetric between the hemispheres.
- Figure 5.
Graph representation of speech-EEG coherence in the attended and ignored conditions for all individual subjects. Rows represent different individuals. Squares denote speech features (i.e., the envelopes from the ten speech bands; shown in the order of increasing center frequency). Each circle denotes an EEG feature (i.e., a particular EEG band from a particular scalp location). An edge between a speech and EEG feature indicates that the coherence between them meets a threshold of 3 SDs from the mean. Only EEG features with one or more edges that survive the thresholding procedure are shown. Attended graphs exhibit greater number of edges compared to ignored graphs for all but two subjects (see bottom two rows). Additionally, the graph structure is variable across subjects. The top two EEG and speech features that are most informative (as obtained using eigenvector centrality) about an individual’s attentional focus also vary across subjects (rightmost column).
- Figure 6.
Individual differences in the overall magnitude of attentional enhancement of speech-EEG coherences in different EEG bands. Each individual’s attentional boost in coherence is shown (with an individual-specific marker symbol and color) for the delta, theta, and low-gamma EEG bands (i.e., the bands showing significant attention effects in Fig. 3A). The mean and SE across individuals are also indicated in black. Note that the y-axis ranges differ between the three panels of the figure. The attentional boost was computed as the percentage change in squared coherence going from the ignored condition to the attended, averaged across EEG channels, speech bands, and the different speech stories. The distribution of the attentional boost across individuals is skewed above zero in all three EEG bands, consistent with positive attentional boost in the neural coding of target speech. Furthermore, there is considerable variation across subjects almost uniformly over the range of boosts.
- Figure 7.
Percentage of edges (i.e., coherences meeting threshold) in attended (ATT) and ignored (IGN) speech-EEG bipartite graphs, at different coherence thresholds. The across-subject distribution of the percentage of graph edges is shown as a violin plot, separately for the attended and ignored conditions, and for three different coherence thresholds. In addition, the median (white dot), 50% confidence limits (thick black box), and 95% confidence limits (black whiskers) of each distribution are shown. Across all three threshold values, the number of edges is significantly larger for the attended condition (based on a permutation test; p values are shown). While Figure 3 showed that specific speech-EEG associations are strengthened by attention, the present result suggests that a greater number of distinct speech-EEG associations are induced by attention.
In this issue
