MyoVoice: A Low-Cost Facial EMG Wearable for Nonverbal Communication

Here is the video for MyoVoice!

Why?

Communication is a fundamental human right, yet over 440,000 Canadians live with significant speech, language, and communication disabilities caused by conditions like cerebral palsy, ALS, Parkinson’s, and stroke. The devices meant to help them cost upwards of $15,000, require fine motor control, or rely on eye-tracking hardware. There are few affordable, non-invasive solution on the market, putting a voice out of reach for those who most need it.

Motivation

When my grandmother had a stroke in early 2024, I saw firsthand the struggle of not having a voice. All she needed was a way to communicate. So, I decided to build MyoVoice to help her and anyone going through the same thing.

Research Gap

Can facial muscles, captured non-invasively, restore real-time communication for individuals with severe speech impairments?

Solution

MyoVoice is a wearable, non-invasive, and super low-cost EMG-based system that translates facial muscle activity into communication commands in real time. It costs $230, much less than current solutions, and gives people back the most important part of their livelihoods.

Goal

An affordable, non-invasive communication system for anyone. I want it to be accessible so that anybody can use it regardless of income or diagnosis.

Approach

Acquire real-time EMG signals from 3 facial muscles.

Extract anatomy-based features that reflect facial movement patterns.

Train and validate a Random Forest Classifier.

Combine predictions over time to ensure that the outputs are as accurate as possible.

Deploy a web-based AAC interface that is ready for immediate use by anyone, anywhere.

For this project, I designed a completely original, EMG-based facial AAC system from scratch! I introduced three novel algorithms that have never been applied to assistive communication.

Hardware

Sensor Design: Three MyoWare surface EMG sensors were placed on the left and right masseters and the submental region. These regions were anatomically selected for intentional facial movement. I went through 4 prototypes to optimize the placement and signal quality.

Signal Acquisition: Sensors interface with an Arduino Bluno, which samples at 100 Hz and streams the synchronized 3-channel data at 115,200 baud.

Signal Processing

Gesture-Onset Extraction (GOE): I noticed that the raw recordings contain significant noise (e.g., periods, delays, and weak start signals). I developed the GOE to automatically isolate the moment when the user moves, so the model learns only from clean data.

Anatomy-Informed Features: Signals are split into 500-ms windows with 50% overlap. I developed custom features based on how facial muscles work together, designed specifically for AAC command classification.

Machine Learning

Classification: After evaluating multiple models, I went with a Random Forest Classifier. A 300-tree Random Forest classifier can classify 4 commands out of the box and can be easily expanded and customized to suit each user's abilities without sacrificing high accuracy.

Sliding Window Confidence Fusion (SWCF): The SWCF uses a rolling prediction buffer, which requires K consecutive high-confidence outputs before triggering a command. This ensures that each output is accurate.

AAC-Interface: A browser-based interface communicates using WebSocket and provides a clear text-to-speech output. New users can be calibrated in under 180 seconds using the Rapid User Adaptation algorithm I created, which weights new data 3x to enable real-time personalization.

To validate my system, I tested it on 3 different users with no prior experience, each calibrated in under 180 seconds, achieving ~99% accuracy across all sessions.

To ensure that MyoVoice was truly robust, I decided to validate it through five different independent methods. I made sure that trial-aware splitting prevented data leakage by ensuring that no windows from the same trial appeared in both training AND test sets.

Results Evaluation

Classification Performance

The held-out test achieved 100% accuracy. To confirm this was not just due to luck, I ran 10 random data splits, yielding 98.61% 1.01%, which is consistently high regardless of how the data is divided. I also used bootstrapping (a statistical technique in which the model is re-evaluated 1,000 times on randomly resampled data), and the accuracy never dropped below 100%, confirming the result with 95% statistical certainty. I also performed an ROC analysis (which measures how well the model separates each command from the others), achieving a perfect mean AUC of 1.0 across all four commands (Figures 9&10).

Visualizing the Features

I used PCA and t-SNE (van der Maaten & Hinton, 2008), two techniques that take complex, multi-feature data and compress it into a single 2D plot, to visualize all 1,472 data points, coloured by command. The four commands formed completely separate, non-overlapping clusters. In other words, the system could clearly “see” the difference between each facial movement, every single time (Figure 11).

Model Confidence/Shannon Entropy

I measured the model's confidence for each prediction using Shannon Entropy, a measure of uncertainty. A score of 0 means perfectly certain; 2.0 means completely random. MyoVoice averaged just 0.093, meaning it was right on almost every prediction.

Leave-One-Trial-Out

In the strictest possible test, I had the model predict each trial, having never seen a single sample from it. Across all trials, it achieved 98.57% accuracy, confirming that the results hold even under the harshest conditions (Figure 12).

Does it beat chance? (McNemar’s Test)

Compared to a random baseline of 25%, MyoVoice’s improvement was statistically significant (p < 0.001) (McNemar, 1947), meaning there is less than a 0.1% chance the results are due to luck.

Prototype Evaluation

Design

As you can see in the project image, the main design for my innovation is a headband. The reason is that a headband is natural for many people and easy to put on without assistance. It also keeps the sensors consistently positioned against the face, which is important for reliable signal quality.

However, the design was also quite personal, as I thought about my grandmother when designing it. I didn’t want it to look clinical or draw attention. I wanted it to feel like something she would actually choose to wear.

Tested on real people

MyoVoice was tested with three distinct participants (of different genders and ages). The 3 new users were calibrated in under 180 seconds each, achieving a mean accuracy of 99.25% (Figure 8), indicating that the system works across multiple users. However, I was unable to conduct further testing on individuals before CWSF for ethical reasons.

Understanding the Impact

MyoVoice proves that restoring communication does not require surgery, eye-tracking hardware, or a five-figure price tag. At $230, up-to-95% less than existing AAC devices, it directly addresses the two biggest barriers facing the 440,000 Canadians with speech disabilities (Communication Disabilities Access Canada, 2023). Those barriers are cost and accessibility. Current solutions require fine motor control, cost tens of thousands of dollars, and take days to calibrate. MyoVoice requires none of that.

Market Opportunity

The global AAC devices market was valued at $2.09 billion in 2024 and is projected to reach $3.62 billion by 2029, growing at a CAGR of 11.5% (Business Research Company, 2025). This growth is driven by a rising demand for affordable, personalized solutions, which is exactly what MyoVoice delivers. Current market leaders charge upwards of $15,000 per device, leaving most users without access (Waters, 2024). MyoVoice is positioned to disrupt this market by delivering superior accuracy at just a fraction of the cost.

Why MyoVoice Wins

No existing AAC system combines non-invasiveness, sub-$300 cost, under-3-minute calibration and ~99% accuracy. MyoVoice does all four: no fine motor control required, works across ALS, cerebral palsy, stroke, and more, and is accessible regardless of income (Statistics Canada, 2022). We have 3 original algorithms, namely the GOE, SWCF, and RUA, which drive performance that existing commercial systems cannot match at ANY price point. The technology to restore communication already exists. MyoVoice proves that it doesn’t have to cost a fortune.

Future Considerations

MyoVoice currently classifies four commands. Future work will expand the vocabulary to support full sentence-level communication, building on recent advances in facial EMG-based human-computer interaction (Perusquia-Hernandez, 2020). Broader clinical validation across larger and more diverse populations is needed to confirm generalizability, particularly for users with ALS and autism spectrum disorder, where facial muscle activity differs from neurotypical controls (Geng et al., 2024). Integrating AI-powered predictive text generation, similar to recent LLM-based AAC approaches, could significantly increase communication speed (HALO, 2025). Finally, miniaturization and wireless transmission would improve comfort and real-world usability, moving MyoVoice closer to a clinical-grade device.

I would like to express my sincere gratitude to my mentor, Dr. Adrian Chan of the Department of Systems and Computer Engineering at Carleton University, for his invaluable guidance and mentorship throughout this long process. I greatly appreciate the feedback you gave me and the time you took out of your busy schedule to help me debug!

I would also like to thank my friend, Daniel Laferriere, for generously allowing me to use his 3D printer at the last minute, which was important to completing the hardware prototype.

I would like to acknowledge the Ottawa Regional Science Fair team for their valuable feedback and guidance, and for their support in preparing this project for CWSF!

Finally, I would like to thank my parents and my extended family for their unwavering support, patience, and belief in me throughout this project and the many nights it took to get it fully working!

ALS Society of Canada. (2024). About ALS. https://als.ca/what-is-als/about-als/

Brain Canada. (n.d.). ALS (amyotrophic lateral sclerosis) research. https://braincanada.ca/brain-conditions/als

Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324

Communication Disabilities Access Canada. (n.d.). People who have communication disabilities. https://www.cdacanada.com/resources/communication-disabilities/statistics/

Government of Canada. (2023). Data blog on parkinsonism in Canada, including Parkinson disease. https://health-infobase.canada.ca/datalab/parkinson-blog.html

Heart and Stroke Foundation of Canada. (2017). Help wanted: Needs not being met for Canadians living with stroke. https://www.heartandstroke.ca/what-we-do/media-centre/news-releases/help-wanted-needs-not-being-met-for-canadians-living-with-stroke

Joyreal. (2026, February 3). Affordable AAC devices: Cost breakdown & budget-friendly options. https://joyrealtoys.com/blogs/news/affordable-aac-communication-devices-budget-friendly-alternatives

Kapur, A., Kapur, S., & Maes, P. (2018). AlterEgo: A personalized wearable silent speech interface. Proceedings of the 23rd International Conference on Intelligent User Interfaces. https://www.researchgate.net/publication/323669071_AlterEgo_A_Personalized_Wearable_Silent_Speech_Interface

McNemar, Q. (1947). Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, 12(2), 153–157. https://doi.org/10.1007/BF02295996

Oskoui, M., Coutinho, F., Dykeman, J., Jetté, N., & Pringsheim, T. (2020). Cerebral palsy in Canada, 2011–2031: Results of a microsimulation modelling study of epidemiological and cost impacts. Health Promotion and Chronic Disease Prevention in Canada, 40(2). https://www.canada.ca/en/public-health/services/reports-publications/health-promotion-chronic-disease-prevention-canada-research-policy-practice/vol-40-no-2-2020/original-quantitative-research-cerebral-palsy-canada-2011-2031.html

Perusquia-Hernandez, M., Dollack, F., Kadone, H., & Suzuki, K. (2020). Facial movement synergies and action unit detection from distal wearable electromyography and computer vision. https://arxiv.org/pdf/2008.08791

Statistics Canada. (2023). Barriers to accessibility related to communication: Findings from the 2022 Canadian Survey on Disability. https://www150.statcan.gc.ca/n1/pub/89-654-x/89-654-x2025003-eng.htm

Statistics Canada. (2023). New data on disability in Canada, 2022. https://www150.statcan.gc.ca/n1/pub/11-627-m/11-627-m2023063-eng.htm

The Business Research Company. (2025). Augmentative and alternative communication (AAC) devices global market report 2025. https://www.thebusinessresearchcompany.com/market-insights/augmentative-and-alternative-communication-aac-devices-market-overview-2025

van der Maaten, L. J. P., & Hinton, G. E. (2008). Visualizing high-dimensional data using t-SNE. Journal of Machine Learning Research, 9, 2579–2605. https://www.jmlr.org/papers/v9/vandermaaten08a.html

Wijdicks, E. F. M. (2014). Adults with cerebral palsy. CMAJ, 197(32), E1034. https://www.cmaj.ca/content/197/32/E1034

-------

Advancer Technologies. (2015). MyoWare muscle sensor (AT-04-001) datasheet. https://cdn.sparkfun.com/datasheets/Sensors/Biometric/MyowareUserManualAT-04-001.pdf

Kohavi, R. (1995). A study of cross-validation and bootstrap for accuracy estimation and model selection. Proceedings of the 14th International Joint Conference on Artificial Intelligence, 2, 1137–1143. https://dl.acm.org/doi/10.5555/1643031.1643047

MyoWare. (2022). MyoWare 2.0 muscle sensor. https://myoware.com/products/muscle-sensor/

Phinyomark, A., Phukpattaranont, P., & Limsakul, C. (2012). Feature reduction and selection for EMG signal classification. Expert Systems with Applications, 39(8), 7420–7431. https://doi.org/10.1016/j.eswa.2012.01.102