Going Beyond Rote Auditory Learning: Neural Patterns of Generalized Auditory Learning

Abstract

The ability to generalize across specific experiences is vital for the recognition of new patterns, especially in speech perception considering acoustic-phonetic pattern variability. Indeed, behavioral research has demonstrated that listeners are able via a process of generalized learning to leverage their experiences of past words said by difficult-to-understand talker to improve their understanding for new words said by that talker. Here, we examine differences in neural responses to generalized versus rote learning in auditory cortical processing by training listeners to understand a novel synthetic talker. Using a pretest-posttest design with EEG, participants were trained using either (1) a large inventory of words where no words were repeated across the experiment (generalized learning) or (2) a small inventory of words where words were repeated (rote learning). Analysis of long-latency auditory evoked potentials at pretest and posttest revealed that rote and generalized learning both produced rapid changes in auditory processing, yet the nature of these changes differed. Generalized learning was marked by an amplitude reduction in the N1-P2 complex and by the presence of a late negativity wave in the auditory evoked potential following training; rote learning was marked only by temporally later scalp topography differences. The early N1-P2 change, found only for generalized learning, is consistent with an active processing account of speech perception, which proposes that the ability to rapidly adjust to the specific vocal characteristics of a new talker (for which rote learning is rare) relies on attentional mechanisms to selectively modify early auditory processing sensitivity.

Figure 1.

Schematic of hypothesized time of changes to the ERP signal as a consequence of generalized learning compared with rote learning. (A) We hypothesize that an overall decrease in potentiation for the longer-latency N1 source or N1c will be observed in the generalized learning condition, but not in the rote learning condition, if the longer-latency N1 source or N1c is sensitive to the demands of attention toward features comprising an auditory object. (B) We hypothesize that an overall decrease in P2 potentiation will be observed in the generalized learning condition, but for not the rote learning condition, if P2 is sensitive to the number of active featural relationships searing current recognition. (C) We hypothesize that an overall decrease in late negativity should be found in the generalized learning condition, but not in the rote learning condition, if late negativity is reflective of a prediction error correction process that supports perception.

Figure 2.

Schematic for trial structure for the generalized learning condition and rote learning condition at pretest, training, and posttest. Trial time is relativized against initial word onset (black bolded screen). All trials start with a fixation cross (1250 msec before initial word onset), followed by a blank screen (500 msec before initial word onset). At 1000 msec after initial word onset, participants are asked to type the word that they heard. In training, this identification procedure was followed by visual written feedback in tandem with an additional auditory presentation of the initial word (gray bolded screen). Participants in the rote condition were given an additional test of 100 novel words (not shown) to behaviorally assess their generalization performance (EEG during this additional test was not recorded).

Figure 3.

Grand-averaged ERPs for generalized learning (A) and rote learning (C) for the nine centralized locations using a virtual montage of the 10–20 system available in BESA Research 7.0 (F3, Fz, F4, C3, Cz, C4, P3, Pz, P4). Grand-average ERPs for P9 and P10 electrodes for generalized learning (B) and rote learning (D) demonstrate the inversion of the N1–P2 complex that is typical of auditory evoked potentials. The ERPs associated with pretest trials are shown in red, and the ERPs associated with posttest trials are shown in blue. Horizontal tick marks span 100 msec, and vertical tick marks represent 1 μV; negative is plotting up. Word onset was used to register and align the EEG traces for averaging, and thus, 0 msec in these plots represents word onset time. Average duration of words in the generalized learning condition was ∼340 msec, whereas the average duration of words in the rote learning condition was ∼350 msec.

Figure 4.

The top plot shows mean GFP (right y-axis) over time for both pretest (short dashed line) and posttest (long dashed line; error bars show ±1 SE) in the generalized learning condition overlaid on the probability over time that the GFP difference under the null was larger than the observed difference in GFP (black line; left y-axis). Significant time periods identified by the GFP analysis are shaded gray: one occurring from 116 to 208 msec and another occurring from 580 to 800 msec. The bottom plot shows the difference in GFP between pretest and posttest (dark black line). For context, the mean (light gray line) and 95% CI (dark gray area) for the GFP difference expected due to random chance (estimated from randomizing the data 5000 times) has been plotted. Significant time periods are again shaded gray, although note that these periods are identified by the observed difference in GFP exceeding the upper bound of the 95% CI of the shuffled data.

Figure 5.

The top plot shows mean GFP (right y-axis) over time for both pretest (short dashed line) and posttest (long dashed line) in the rote learning condition (error bars show ±1 SE) overlaid on the probability over time that the GFP difference under the null was larger than the observed difference in GFP (black line; left y-axis). No significant time windows were observed. The bottom plot shows the difference in GFP between pretest and posttest. For context, the mean (light gray line) and 95% CI (dark gray area) for the GFP difference expected due to random chance (estimated from randomizing the data 5000 times) has been plotted.

Figure 6.

Plot A shows how the generalized dissimilarity statistic between pretest and posttest topographies varies overtime in the generalized learning condition (black line). For context, the mean (light gray line) and 95% CI (medium gray area) for the generalized dissimilarity statistic expected due to random chance (estimated from randomizing the data 5000 times) has been plotted. Plot W1 shows the results of the spatiotemporal permutation-based analysis that was performed on the W1 window found in RAGU. This plot shows the average topographic difference between pretest and posttest (contrast: posttest–pretest) and indicates where the three significant electrode clusters (p < .05) are topographically located.

Figure 7.

Plot A shows the time-varying generalized dissimilarity between pretest and posttest topographies for rote learning (black line). For context, the mean (light gray line) and 95% CI (medium gray area) for the generalized dissimilarity expected due to random chance (estimated from randomizing the data 5000 times) has been plotted. Six windows were identified where the observed data exceeded the upper bound of the 95% CI of the shuffled data. Plots W1, W2, W3, W4, W5, and W6 show the results of the spatiotemporal permutation-based analysis that was performed on each window. All electrode clusters shown, uniquely colored on each plot, are significant at a p < .05 level.