Dynamics of self-monitoring and error detection in speech production: evidence from mental imagery and MEG

Abstract

A critical subroutine of self-monitoring during speech production is to detect any deviance between expected and actual auditory feedback. Here we investigated the associated neural dynamics using MEG recording in mental-imagery-of-speech paradigms. Participants covertly articulated the vowel /a/; their own (individually recorded) speech was played back, with parametric manipulation using four levels of pitch shift, crossed with four levels of onset delay. A nonmonotonic function was observed in early auditory responses when the onset delay was shorter than 100 msec: Suppression was observed for normal playback, but enhancement for pitch-shifted playback; however, the magnitude of enhancement decreased at the largest level of pitch shift that was out of pitch range for normal conversion, as suggested in two behavioral experiments. No difference was observed among different types of playback when the onset delay was longer than 100 msec. These results suggest that the prediction suppresses the response to normal feedback, which mediates source monitoring. When auditory feedback does not match the prediction, an "error term" is generated, which underlies deviance detection. We argue that, based on the observed nonmonotonic function, a frequency window (addressing spectral difference) and a time window (constraining temporal difference) jointly regulate the comparison between prediction and feedback in speech.

Figure 1

Proposed model: Kalman gain modulates errors terms (e_f) between internal prediction and external feedback. The Kalman gain (K; Equation 1), which is a function of spectral (D) and temporal (T) differences between prediction and feedback, actively modulates the magnitude of error terms that are formed by comparing the prediction (P) and feedback (F).

Figure 2

Schematic description of the experimental design. Participants press a button at the beginning of articulation imagery of the vowel /a/. The prerecorded individual vocalization of /a/ was manipulated and formed four levels of pitch-shifted playback that was presented at four levels of delays after the button press.

Figure 3

Behavioral results of pitch range assessment. The subjective probability judgment of pitch range in one's normal conversation is plotted as a function of pitch shift. Left: Results in Behavioral Experiment 1 using probability judgment. Right: Results in Behavioral Experiment 2 using categorization task. The red dash lines represented the chance level.

Figure 4

Waveform responses to playback in all conditions during experimental and baseline runs. RMS waveform responses to auditory feedback in all conditions were plotted. Two lines were included in each subplot, with the red line representing the response during experimental run and the black line for baseline run. Subplots were arranged vertically as the level of pitch shift increases and horizontally as the level of onset delay increases.

Figure 5

Topographies of auditory M100 responses to playback in all conditions during experimental and baseline runs. Topographies are grouped in a pair for each condition, with the one in experimental run on left and baseline run on right. The layout of topographies for all conditions is identical as Figure 4.

Figure 6

Response differences in M100 (left) and M200 (right) components as functions of pitch shift and onset delay. The relative magnitude changes, calculated by subtracting the auditory response in baseline run from the one in experimental run, are plotted as a function of pitch shift and grouped by each level of onset delay.

Figure 7

Waveform responses and topographies of control runs. Left: Waveform responses and topographies of four different pitch shifted sounds in auditory control. RMS waveforms to four pitch shifted sounds are plotted in different colors. M100 and M200 components were inserted beside the waveforms at corresponding latencies. Color boxes that indicate topography of responses to each sound use the same color code as waveforms. Right: Waveforms responses and topographies of button press in motor control runs. RMS waveforms to button press responses are plotted for runs with (red) and without (black) articulation imagery. Two response components, presumably mediating button press and release, are plotted near the corresponding latency, with the surrounding boxes being identical color code as waveforms.