To map how the human brain processes sound or how the vocal tract shapes speech, researchers rely heavily on Magnetic Resonance Imaging (MRI). However, the very physics that enable MRI—the rapid switching of magnetic field gradients—generate intense, high-decibel mechanical vibrations and acoustic noise.
This creates a fundamental paradox: the primary instrument used to measure auditory processing acts as a massive, uncontrolled auditory stimulus itself. This acoustic scanner noise degrades the signal-to-noise ratio (SNR) of both neural data and concurrent audio recordings, presenting a pervasive challenge across cognitive neuroscience, clinical audiology, and speech science.
How Scanner Noise Contaminates Neural Signals
In functional neuroimaging (fMRI), the scanner’s ambient noise acts as an environmental masker and cognitive distractor, fundamentally altering the neural landscape under investigation.
Auditory Masking and Distraction
The continuous or pulsed roar of the scanner can obscure subtle neural responses to experimental stimuli. As noted by Irene Neuner et al. in their study, this background noise acts as a significant confounder that interferes with the neural processing of target auditory stimuli. It serves as an uncontrolled auditory distractor, which Andrew R. Mayer et al. found complicates the isolation of neural activity related to sensory tuning and attentional filtering in their research.
Cognitive Overload and Altered Processing
When participants attempt to process speech over this background roar, their brains must expend extra cognitive effort. Patrick C.M. Wong et al. demonstrated in their work that listening to speech in noise significantly increases left-hemisphere activity, suggesting that scanner noise forces the brain to recruit compensatory mechanisms like subvocal rehearsal.
Similarly, Stanislas Dehaene pointed out in his study that the noisy environment can interfere with top-down attentional processes. This acoustic baseline shift can even weaken the detection of subtle behavioral effects; for instance, Sid Kouider et al. observed in their paper that scanner noise attenuated the behavioral effects of subliminal speech priming.
Baseline Contamination in Resting-State fMRI
The issue becomes even more acute in resting-state fMRI (rs-fMRI), where researchers attempt to measure spontaneous brain activity in the absence of external tasks. Because the scanner noise is persistent, it serves as an unintended external auditory stimulus that drives neural activity throughout the scan.
Xiaoxiao Xie et al. showed in their study that this contamination is particularly problematic when mapping the functional networks of patients with hearing impairments. Furthermore, Yu‐Chen Chen et al. noted in their work that this ambient noise distorts measurements of spontaneous brain activity in patients with phantom sound perceptions like tinnitus, making it difficult to isolate true baseline connectivity.
(Note: Similar noise artifacts can occur in other magnetic modalities; for example, Xiayang Tian et al. demonstrated in their paper that high currents required for magnetic neuromodulation generate audible noise and vibrations that can confound animal behavioral experiments independently of the intended magnetic effects).
Disrupting Speech Production and Articulatory Mapping
Beyond auditory perception, scanner noise severely impacts studies of speech production. Real-time MRI (rtMRI) allows scientists to capture high-speed videos of the vocal tract, tongue, and throat in motion. However, the accompanying mechanical noise makes concurrent audio recordings nearly unusable.
[Vocal Tract Motion] ---> Captured perfectly by rtMRI
|
v
[Acoustic Environment] -> Contaminated by ~100 dB Scanner Noise
|
v
[Audio Recording] ------> Severely degraded / Unusable for synthesis
Ruined Audio and Feedback Loops
As Ioannis Douros highlighted when constructing a French articulatory corpus in their paper, high levels of acoustic noise frequently obscure the speaker's voice. This degradation prevents researchers from directly correlating physical articulator movements with clean, synchronized speech sounds—a challenge also emphasized by Xiaofeng Liu et al. in their cross-modal translation work.
Furthermore, scanner noise alters the speaker's own experience. Tarek A. Yousry argued in his study that scanner noise acts as a major confound during speech production because it obscures the normal auditory feedback loops that speakers rely on to monitor their own voice.
Altered Articulation
The scanner environment itself changes how subjects behave. Gaowu Wang et al. found in their research that the high-noise and claustrophobic environment induces subject fatigue and alters natural articulation. This can lead to discrepancies in phonetic data, particularly for complex phonetic sequences that require high stability or physical effort.
Mitigation Strategies and Algorithmic Solutions
To overcome these limitations, researchers have developed both hardware-level temporal adjustments and software-level algorithmic workarounds.
Optimizing Acquisition Timing
One classic approach to mitigating noise in fMRI is to manipulate the sequence repeat time (TR) or introduce silent gaps between scans. Because the scanner only makes noise during data acquisition, researchers can present auditory stimuli during periods of silence.
N. Jon Shah et al. investigated this in their study, demonstrating that adjusting the TR can help balance scanner noise and measurement efficiency, finding an optimal TR of around 6 seconds for phonetic discrimination tasks. Similarly, the authors of a study on natural sound encoding utilized silent gap acquisition to present auditory stimuli in quiet, though they noted that the persistent background context can still distort the neural representation of those sounds.
Bypassing Audio via Machine Learning
When clean audio cannot be recorded due to intense mechanical noise during real-time imaging, modern deep learning offers a way to bypass the acoustic channel entirely. Instead of trying to filter out the noise from a heavily corrupted audio track, recent frameworks attempt to synthesize speech directly from the silent video of the moving vocal tract.
[rtMRI Video] ---> [Predict Text Tokens] ---> [Duration-Aware Decoder] ---> [Clean Speech Synthesis]
In their paper, Neil Kumar Shah et al. introduced MRI2Speech, a framework that shifts from direct video-to-audio mapping to a text-mediated approach. Traditional models struggle because training them requires clean ground-truth audio, which is impossible to collect inside a noisy scanner.
By training a model to first predict text from the silent video frames and then using a duration-aware decoder to synthesize speech from that text, the system completely sidesteps the acoustic noise of the scanner. This architectural workaround achieved a massive improvement in intelligibility, dropping the Word Error Rate (WER) from over 100% in prior direct-mapping methods to just 15.18%.
Key Foundational Work
The Effect of Sequence Repeat Time on Auditory Cortex Stimulation During Phonetic Discrimination (2000) — contributed 4 concept elements to this direction.
Neural bases of listening to speech in noise (2005) — contributed 1 concept elements to this direction.
Neural correlates of switching from auditory to speech perception (2005) — contributed 1 concept elements to this direction.