Figure 4.

(a) The structure of the Generator network with five convolutional layers⁶⁶. The fourth convolutional layer (Conv4; second to last) is color-coded with purple. (b) All 64 individual feature maps for a single output forced to closely resemble the stimulus from the fourth convolutional layer (Conv4) after ReLU (upsampled). (c) One Spanish output (in green) forced to resemble the stimulus with the corresponding values from the fourth convolutional layer (Conv4) averaged over all feature maps. The plot illustrates peak latency between output and Conv4 for the burst and each vocalic period. (d) A zoomed version of (c) focusing on four vocalic periods. (e) One English output (in green) forced to resemble the stimulus with the corresponding values from the fourth convolutional layer (Conv4) averaged over all feature maps. The plot illustrates peak latency between output and Conv4 for the burst and each vocalic period. (f) A zoomed version of (e) focusing on four vocalic periods. (g) Raw peak latency timing (output peak time - Conv4 peak time) for burst (=0) and each nth vocalic period across the two conditions (English vs. Spanish). Periods above the 12th period are rare and are discarded from the statistical analysis due to a small number of attestations. The data is pooled across the two replications. (h) Raw peak latency timing across the replications (first and second replication) and two conditions (English and Spanish). (i) Linear regression estimates for the peak latency timing between the two conditions (English vs. Spanish). Periods above the 12th period are discarded from the analysis due to a small number of attestations. The data is pooled across the two replications.