Abstract
The study of speech planning/programming may require analysing Event-Related Potentials (ERPs) during articulation. However, ERPs identified during speech production also contain brain signals associated with auditory feedback. Because these processes are both time-locked to the vocal onset, existing algorithms for signal separation have difficulties distinguishing one from the other. Here, we investigated the use of the multi-channel Wiener filter (MWF) to remove the ERP signal related to auditory processing from the ERP signal related to speech production. In a first step, participants were asked to overtly produce mono- or disyllabic pseudowords. In a second step, they had to listen to the recording of their own productions. We used the ERP signal associated with hearing one's own production (referred to as listening ERPs) to estimate and filter out the auditory part of the production ERP. We investigated three versions of the ERPs during speech articulation: (1) the original production ERPs, (2) the ERPs filtered through the MWF, and (3) subtraction ERPs obtained by subtracting the listening ERPs from the original ERPs. Firstly, we conducted a microstate analysis comparing all three versions of the ERPs. Secondly, we conducted separate microstate analyses comparing mono- versus disyllabic conditions on each of the three versions. The results indicate that the subtraction method alters topography consistency and may remove relevant signal from the ERPs. The filtered ERP produced similar results to the original production ERP and thus, it may be unnecessary to remove auditory feedback or other overlapping signal if time-locked to the production onset.
Introduction
In neurolinguistics, Event-Related Potentials (ERPs) are essential for tracking brain activity during language tasks. While most studies focus on the "pre-vocal" period—the split second before speech begins—speech planning likely continues during articulation, particularly for multi-syllabic words. However, as soon as vocalization starts, the brain processes the sound of its own voice. This Auditory Evoked Potential (AEP) overlaps perfectly in time with motor and planning signals, creating a significant challenge for signal separation. This study evaluates whether removing this auditory "interference" reveals previously hidden details about speech planning.
Methods
The researchers recorded EEG data from 17 native French speakers during two distinct tasks. In the Delayed Production Task, participants saw a pseudoword (e.g., "klacha"), waited for a cue, and then spoke. In the Listening Task, they heard recordings of their own previous productions. Pseudowords were used to ensure the brain focused on phonological planning rather than semantic retrieval.
Three processing methods were compared using Microstate Analysis, which identifies stable periods of brain activity by looking at the total topography of electrical potential across the scalp:
- Original Production ERPs: The raw, averaged signal.
- Subtraction ERPs: A simple mathematical subtraction of the listening ERP from the production ERP.
- Multi-channel Wiener Filter (MWF): A supervised statistical filtering technique that uses the listening task as a template to suppress auditory noise while preserving the production signal.
Results
The listening task successfully captured the standard AEP components, including the P1 and N1 peaks, which represent the brain's initial response to sound.
Key findings include:
- The Subtraction Fallacy: Simple subtraction was found to be problematic. Because the brain partially inhibits the auditory cortex during actual speech (sensory suppression), the listening task is not a perfect proxy for the auditory feedback during production. Subtraction led to "topographic inconsistency," effectively introducing noise.
- MWF Effectiveness: The MWF proved to be a precise tool, successfully removing signal components associated with the listening template without destroying the underlying structure of the production data.
- Redundancy of Filtering: Despite the technical success of MWF, the filtered data did not reveal any new differences between mono- and disyllabic words that were not already visible in the original signal. This suggests that if experimental conditions are balanced, overlapping auditory noise may cancel itself out during comparison.
Discussion and Conclusion
While sophisticated tools like the Multi-channel Wiener Filter can effectively separate overlapping brain signals, they may be unnecessary for many speech production studies. The study demonstrated that the original production ERPs are robust enough to detect planning differences, such as the increased complexity required for disyllabic words.
Future research should address "temporal jitter"—the possibility that speech planning for subsequent syllables is not perfectly time-locked to the initial vocal onset. In such cases, standard ERP averaging might wash out planning signals regardless of filtering quality. Overall, the findings suggest that for well-balanced experimental designs, the "noise" of hearing oneself does not significantly mask the underlying cognitive-motor processes of speech production.
Summary updated