Inter-Trial Formant Variability in Speech Production Is Actively Controlled but Does Not Affect Subsequent Adaptation to a Predictable Formant Perturbation

Abstract

Despite ample evidence that speech production is associated with extensive trial-to-trial variability, it remains unclear whether this variability represents merely unwanted system noise or an actively regulated mechanism that is fundamental for maintaining and adapting accurate speech movements. Recent work on upper limb movements suggest that inter-trial variability may be not only actively regulated based on sensory feedback, but also provide a type of workspace exploration that facilitates sensorimotor learning. We therefore investigated whether experimentally reducing or magnifying inter-trial formant variability in the real-time auditory feedback during speech production (a) leads to adjustments in formant production variability that compensate for the manipulation, (b) changes the temporal structure of formant adjustments across productions, and (c) enhances learning in a subsequent adaptation task in which a predictable formant-shift perturbation is applied to the feedback signal. Results show that subjects gradually increased formant variability in their productions when hearing auditory feedback with reduced variability, but subsequent formant-shift adaptation was not affected by either reducing or magnifying the perceived variability. Thus, findings provide evidence for speakers' active control of inter-trial formant variability based on auditory feedback from previous trials, but-at least for the current short-term experimental manipulation of feedback variability-not for a role of this variability regulation mechanism in subsequent auditory-motor learning.

Keywords: acoustics; adaptation; articulation; auditory feedback; speech motor control; variability.

FIGURE 1

(A) Speech tasks completed by two groups of participants. Within each group, order of the experimental condition (Magnified or Attenuated feedback variability) and the Control condition was counterbalanced across participants. (B) Instrumentation setup. (C) Example spectrograms of Difference-shifted trials in the Magnified and Attenuated conditions of the Variability task. Dashed yellow line: pre-test median formant frequencies (F1 and F2, in Hz). Solid magenta and orange lines: produced (left of arrow) and heard (right of arrow) formants in the Magnified and Attenuated conditions. (D) Time course of the formant-shift feedback perturbation in the Adaptation task.

FIGURE 2

(A,C) Example individual participant data for production and feedback Distance Index (DI) of each trial in the Variability task under Magnified and Attenuated conditions. (B,D) Boxplots with symbols depicting each participant’s ratio between average feedback DI and average production DI (both measured mid-vowel) across all trials of the Variability task with Magnified or Attenuated feedback variability.

FIGURE 3

(A,C) Boxplots with symbols depicting each participant’s average DI_initial for the entire Variability task in Control and Magnified or Control and Attenuated conditions. (B,D) Boxplots with each participant’s average DI_mid for the entire Variability task in the same conditions.

FIGURE 4

(A,B) Change in DI_initial across the Variability task by block (i.e., 3 trials) for the Magnified and Attenuated feedback variability conditions. Dots represent the group mean DI per block. Shaded regions indicate standard error of the mean (SEM). Solid lines are loess smoothed fits (span = 0.6). (C,D) Change in DI_initial across the Variability task by stage (i.e., 15 trials) for the Magnified and Attenuated feedback variability conditions. Error bars indicate SEM. Asterisks indicate adjusted p < 0.05 (see Table 1). (E–H) Individual participant data for the significant change from Stage 1 to Stage 3 in the Attenuated condition: (E) Stripchart of DI_initial in Stage 1 and Stage 3. Horizontal lines indicate deciles; bold line is the median. (F) Stripchart with each participant’s Stage 1 and Stage 3 data linked. (G) Scatterplot of Stage 1 by Stage 3 data. The diagonal line denotes no difference between stages. Participants in the upper left half increased DI_initial in Stage 3. Dashed lines mark quartiles. (H) Stripchart of the difference in DI_initial between Stage 3 and Stage 1. Horizontal lines indicate deciles; the bold line is the median; the dashed line is at zero (no difference between stages).

FIGURE 5

(A,B) Change in DI_mid across the Variability task by block (i.e., 3 trials) for the Magnified and Attenuated feedback variability conditions. Dots represent the group mean DI per block. Shaded regions indicate standard error of the mean (SEM). Solid lines are loess smoothed fits (span = 0.6). (C,D) Change in DI_mid across the Variability task by stage (i.e., 15 trials) for the Magnified and Attenuated feedback variability conditions. Error bars indicate SEM. Asterisks indicate adjusted p < 0.05 (see Table 1). (E–H) Individual participant data for the significant change from Stage 1 to Stage 3 in the Attenuated condition: (E) Stripchart of DI_mid in Stage 1 and Stage 3. Horizontal lines indicate deciles; bold line is the median. (F) Stripchart with each participant’s Stage 1 and Stage 3 data linked. (G) Scatterplot of Stage 1 by Stage 3 data. The diagonal line denotes no difference between stages. Participants in the upper left half increased DI_mid in Stage 3. Dashed lines mark quartiles. (H) Stripchart of the difference in DI_mid between Stage 3 and Stage 1. Horizontal lines indicate deciles; the bold line is the median; the dashed line is at zero (no difference between stages).

FIGURE 6

Individual participant data (one participant per panel) for inter-trial formant dispersion in acoustic vowel space (F1 by F2). Data based on 95% confidence ellipses, calculated for formant frequencies extracted from the initial portion of the vowels. Each participant’s data from Stages 1 and 3 (15 trials per stage) in the Attenuated feedback variability condition are shown together with their data from the Pre-test (90 trials). Nine of 13 participants increased ellipse area in Stage 3 as compared with Stage 1. Participants are ordered (by row) from greatest to smallest ellipse area increase.

FIGURE 7

Sample lag 1 autocorrelation functions [ACF(1)] for formant data measured in the initial portion of the vowel and averaged across F1 and F2 for Control versus Magnified (A) and Control versus Attenuated (B) conditions of the Variability task. Dashed lines indicate the large sample 95% confidence interval of ACF(1) for a white noise process with sample size 75 (the number of trials per condition). Each dot represents an individual participant.

FIGURE 8

Group-level formant-shift adaptation data after completion of the Variability task’s Control and Magnified conditions (A) or after the Control and Attenuated conditions (B). Dots represent group mean formant frequencies per block (3 trials) and averaged across F1 and F2. Shaded regions indicate standard error of the mean. Solid lines are loess smoothed fits (span = 0.3).

FIGURE 9

Boxplots with symbols depicting each participant’s early adaptation extent (A,B), early adaptation rate (C,D), and final adaptation extent (E,F) for formant-shift adaptation completed after the Variability task’s Control and Magnified or Control and Attenuated conditions. Full statistics for these data: (A) t(13) = 0.366, p = 0.720, d = 0.098; (B) t(12) = –0.822, p = 0.427, d = 0.228; (C) t(13) = 0.150, p = 0.883, d = 0.040; (D) t(13) = –0.128, p = 0.900, d = 0.035; (E) t(13) = 0.301, p = 0.768, d = 0.081; (F) t(12) = –0.450, p = 0.661, d = 0.125.