⚡ electrical engineering

Efficient Sensor Configuration for Accelerometer-based Silent Speech Interfaces: Regionally Distributed Sensor Placement Strategy Considering Articulatory Dynamics

This study identifies an optimal five-sensor configuration spanning diverse functional articulatory regions on the lower face and neck that achieves 92.30% accuracy in silent speech classification, demonstrating that regionally distributed sensor placement is a superior design strategy for compact accelerometer-based silent speech interfaces.

Original authors: Heejin Choi, Sungmin Jung, Jangjay Sohn, Chang-Hwan Im

Published 2026-07-23

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Heejin Choi, Sungmin Jung, Jangjay Sohn, Chang-Hwan Im

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Silent Whisper: Catching Words Without Sound

Imagine trying to have a conversation in a library so quiet that even a whisper would echo, or perhaps you are in a situation where speaking aloud is impossible—maybe you've lost your voice, or you're in a noisy factory where no one can hear you. For centuries, humans have relied on sound waves to share thoughts, but what if we could talk without making a single sound? This is the world of Silent Speech Interfaces (SSIs). Think of them as a secret decoder ring for your mouth. Instead of listening to your voice, these devices listen to the tiny, invisible movements your lips, jaw, and throat make when you "speak" silently.

To do this, scientists use different tools. Some use cameras to watch your lips move, while others use sensors that feel the electrical zaps from your muscles. But there's a newer, cooler tool: the accelerometer. You might know these from your smartphone, which knows when you tilt your phone or shake it. In this research, scientists strap tiny accelerometers to your face and neck. These sensors don't care about electricity or cameras; they only care about motion. When you silently say "apple," your chin drops, your lips pucker, and your throat shifts. The accelerometer feels that tiny bump and records it as a wiggly line of data. The big question, however, is: Where should you put these sensors? If you put them all in one spot, you might miss the action. If you put too many, the device becomes a heavy, uncomfortable mess. Finding the perfect spot is like trying to find the best seat in a theater to see the whole show without getting a crick in your neck.

The Great Sensor Hunt: Finding the Sweet Spot

In this study, a team of researchers from Hanyang University decided to play a massive game of "Where's Waldo?" but with sensors on a human face. They wanted to figure out the most efficient way to strap accelerometers to a person's lower face and neck to decode 100 different English words spoken silently. They didn't just guess; they tested every single possible combination of ten sensors. Imagine having ten different friends, and you want to know which group of friends works best together to solve a puzzle. You could try groups of one, groups of two, all the way up to all ten. That's exactly what they did, testing over 1,000 different sensor arrangements with 20 volunteers.

The volunteers sat in front of a screen and silently mouthed 100 words, like "hello," "stop," or "blue," while wearing the sensors. A smart computer program (a type of artificial intelligence called a deep learning model) then tried to guess which word was being said based on the wiggly lines from the sensors.

Here is what they found:

1. More isn't always better (The Law of Diminishing Returns)
At first, adding more sensors helped the computer guess better. With just one sensor, the computer was right about 75% of the time. With two sensors, it jumped to over 80%. But here is the twist: once they reached five sensors, adding more didn't really help. The accuracy hit a "ceiling" or a plateau. Whether they used six, seven, or all ten sensors, the improvement was so tiny it wasn't even statistically significant. It's like adding more chefs to a kitchen; after a certain point, they just get in each other's way without making the meal taste better.

2. The Golden Five
The researchers found a specific group of five sensors that worked the best. This "dream team" consisted of sensors placed at:

The Philtrum (the little dip between your nose and upper lip).
The Lower Lip/Chin area (two spots).
The Jawline (two spots).
The Under-chin/Throat area.

This specific mix, labeled as sensors {1, 4, 5, 7, 8}, achieved an accuracy of 92.30 ± 4.64%. This was almost as good as using all ten sensors combined, but with half the hardware.

3. The "Spread Out" Strategy
The most exciting discovery wasn't just how many sensors they used, but where they were. The researchers grouped the face into three "zones" of movement:

Labial (Lips): Where the lips move.
Buccal-Mandibular (Jaw/Cheek): Where the jaw opens and closes.
Submental (Under the chin/Throat): Where the throat and tongue move.

They discovered that the best sensor setups were the ones that spread out across all three zones. A setup that put all five sensors on just the lips (ignoring the jaw and throat) did much worse than a setup that put one sensor on the lips, two on the jaw, and two on the throat. It's like trying to understand a movie by only watching the actors' hands; you miss the facial expressions and the body language. To understand the "silent speech," you need to see the whole picture. The best configurations captured the "complementary" information—meaning the sensors worked together to fill in the gaps left by the others.

4. Does it work with different words?
To make sure this wasn't a fluke, the team tried the experiment again with a different list of 100 words. The exact sensors changed slightly (one sensor on the jaw was swapped for one on the neck), but the strategy stayed the same: the best results always came from spreading sensors across the lips, jaw, and throat. This suggests that the "spread out" rule is a solid guide for building these devices, no matter what words you want to say.

5. The "Mask" Idea
Finally, the team asked: "Can we make this even simpler?" They tried to find a setup with just three sensors that still covered all three zones. They picked one sensor from the lips, one from the jaw, and one from the throat. This tiny trio still managed to guess words with nearly 90% accuracy. This is a huge deal because it suggests that in the future, we might be able to wear a simple, comfortable mask or a small patch on the neck that can decode our silent thoughts without needing a dozen wires and sensors stuck all over our face.

What This Means for the Future

This paper didn't just find a magic sensor; it changed the way we think about building these devices. Before this, scientists often just copied where they put sensors from other types of technology (like muscle sensors) without asking if it was the best spot for motion sensors. This study proved that for accelerometers, location matters more than quantity.

The researchers suggest that instead of piling on sensors, we should design them to be "regionally distributed." Think of it like a security team: you don't put all your guards in the front door; you put some at the back, some at the windows, and some in the lobby. By spreading the sensors across the different moving parts of the mouth and throat, the device gets a complete 3D picture of the silent speech.

While the study was done with specific people and specific words, the results strongly suggest that this "spread out" strategy is the key to making silent speech devices that are small, comfortable, and accurate enough for real life. Whether for people who have lost their voices or for secret agents who need to talk without being heard, the future of silent speech might just be a few tiny sensors placed in the right spots.

Technical Summary: Efficient Sensor Configuration for Accelerometer-based Silent Speech Interfaces

Problem Statement

Silent Speech Interfaces (SSIs) aim to decode articulatory activity without audible speech, offering critical communication solutions for individuals with laryngectomy and enabling discreet interaction in noisy or secure environments. While various sensing modalities exist (e.g., sEMG, ultrasound, video), accelerometers offer a compact, durable, and lightweight alternative for capturing facial and neck movements. However, a significant gap exists in the systematic optimization of sensor placement for accelerometer-based SSIs.

Previous research often inherited sensor locations from sEMG-based studies, which measure electrical muscle activity, whereas accelerometers measure physical motion. This mismatch can lead to suboptimal configurations, as demonstrated by prior work where excluding the zygomatic site (effective for sEMG) improved accelerometer performance. Furthermore, increasing the number of sensors increases user discomfort and setup complexity. There is currently a lack of evidence regarding the optimal number of sensors and the strategic distribution of these sensors across functional articulatory regions to maximize classification accuracy while minimizing hardware burden.

Methodology

Data Acquisition and Experimental Design

Participants: 20 native Korean-speaking adults (10 male, 10 female) familiar with English.
Task: Classification of 100 silently spoken English words.
Hardware: Ten 3-axis accelerometers (MPU-9250) were distributed across the lower face and neck.
Sensor Grouping: The ten locations were categorized into three functional articulatory regions based on kinematic characteristics:
1. Labial Region: Sensors near the lips (Sensors 1, 2, 7).
2. Buccal-Mandibular Region: Sensors along the jawline and cheek (Sensors 3, 4, 6, 8).
3. Submental Region: Sensors on the neck and upper throat (Sensors 5, 9, 10).
Procedure: Participants silently articulated words within a 4-second window across five sessions. Data was recorded at 100 Hz, downsampled to 50 Hz, and normalized.

Model Architecture

A deep learning model consisting of a Convolutional Neural Network (CNN) followed by a Bidirectional Long Short-Term Memory (BiLSTM) network was employed. The CNN extracted local temporal features, and the BiLSTM modeled temporal dependencies. The model was trained using subject-dependent cross-validation (10-fold) with Adam optimization.

Analysis Strategy

The study systematically evaluated sensor configurations through two primary lenses:

Sensor Count Optimization: All 1,023 possible combinations of the ten sensors were evaluated to identify the "saturation point" where adding more sensors yields no statistically significant accuracy gain.
Regional Coverage Analysis: Configurations were compared based on the number of functional regions they covered (single-region vs. multi-region) to determine if distributed placement offers advantages over localized clustering.
Validation: The findings were validated using a second, distinct 100-word vocabulary set to assess generalizability.

Key Results

Optimal Sensor Count and Configuration

Saturation Point: Classification accuracy increased with the number of sensors but plateaued at five sensors. Adding more than five sensors did not result in a statistically significant improvement in accuracy ( $p > 0.05$ ).
Efficient Configuration: The five-sensor configuration {1, 4, 5, 7, 8} achieved the highest accuracy of 92.30 ± 4.64%. This configuration showed no significant difference from the overall best-performing configuration (which used eight sensors).
Sensor Locations: The efficient configuration includes:
- Labial: Sensors 1 (philtrum) and 7 (below lower lip).
- Buccal-Mandibular: Sensors 4 and 8 (jawline/cheek).
- Submental: Sensor 5 (neck).

Impact of Regional Coverage

Multi-Region Superiority: Configurations spanning multiple functional articulatory regions consistently outperformed those restricted to a single region.
- In a fixed 3-sensor comparison, multi-region configurations significantly outperformed single-region ones ( $p < 0.001$ ).
- At the 5-sensor saturation point, the full-coverage configuration ({1, 4, 5, 7, 8}) significantly outperformed the best partial-coverage configuration ( $p = 0.0009$ ).
Complementary Information: Ablation and perturbation analyses revealed that the selected sensors provided unique, non-redundant information. For instance, removing the submental sensor (Sensor 5) caused a larger accuracy drop than removing buccal-mandibular sensors, suggesting that tongue/vocal-tract related motion captured by the neck is not fully represented by facial sensors.

Generalizability and Practicality

Vocabulary Independence: When tested on a new 100-word vocabulary, the performance saturation pattern remained similar. While the specific optimal sensor indices shifted slightly (e.g., Sensor 4 was replaced by Sensor 10), the highest-performing configurations consistently maintained coverage across all three functional regions.
Reduced Configuration: A practical 3-sensor candidate {5, 7, 8} (covering all three regions) achieved nearly 90% accuracy, demonstrating the feasibility of a mask-type wearable device.

Key Contributions and Significance

The authors claim this study provides the first systematic investigation into both sensor count and placement strategy specifically for accelerometer-based SSIs. The primary contributions are:

Shift in Design Focus: Moving the research focus from purely algorithmic improvements (e.g., better deep learning models or data augmentation) to the fundamental hardware design factor of sensor placement.
Accelerometer-Specific Evidence: Demonstrating that sensor placements effective for sEMG are not necessarily optimal for accelerometers due to the difference between measuring electrical potential versus physical motion.
Regionally Distributed Placement Strategy: Proposing a design principle that prioritizes distributing sensors across distinct functional articulatory regions (labial, buccal-mandibular, submental) to capture complementary articulatory information, rather than clustering sensors in a single area.
Efficiency Benchmark: Identifying that five sensors are sufficient to reach performance saturation, offering a concrete guideline for reducing hardware complexity and user burden without sacrificing recognition accuracy.

The paper concludes that this regionally distributed strategy serves as a robust design basis for future compact, wearable accelerometer-based SSI systems, validated across different vocabularies and capable of supporting real-time classification.