Silent and Vocalized Speech Classification Using Surface Electromyography and Machine Learning
Resumo
Communication impairments arising from medical conditions such as laryngectomies, tracheostomies, or neurological disorders significantly limit the quality of life of affected individuals. Surface electromyography (EMG) offers a promising alternative for speech recognition by capturing muscle activity without requiring audible sound production. This paper investigates the classification of silent and vocalized speech commands using EMG signals and machine learning algorithms, with the hypothesis that discriminative time-domain features are sufficient to enable effective classification even under silent speech conditions. Thirteen features were extracted from the “SilentWear” dataset, which contains recordings of eight HMI commands from four subjects, using an event-based approach across 14 EMG channels, yielding 182-dimensional feature vectors. Various dimensionality reduction techniques were applied to the feature space, and three classification algorithms were evaluated on them using Leave-One-Subject-Out (LOSO) and Leave-One-Trial-Out (LOTO) crossvalidation schemes across multiple experimental scenarios. LDA achieved the best overall performance, reaching an accuracy of 0.829 ± 0.075 and a Macro-F1 of 0.664± 0.151 for vocalized speech under LOTO, and demonstrating competitive results even for silent speech, which highlights the discriminative power of the adopted feature set.
Referências
Afshar, S., Ralph, N., Xu, Y., Tapson, J., van Schaik, A., and Cohen, G. (2020). Event-based feature extraction using adaptive selection thresholds. Sensors, 20(6):1600.
Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer, New York.
Breiman, L. (2001). Random forests. Machine Learning, 45(1):5–32.
Cortes, C. and Vapnik, V. (1995). Support-vector networks. Machine Learning, 20(3):273–297.
Dellacasa Bellingegni, A., Gruppioni, E., Colazzo, G., Davalli, A., Sacchetti, R., Guglielmelli, E., and Zollo, L. (2017). Nlr, mlp, svm, and lda: A comparative analysis on emg data from people with trans-radial amputation. Journal of NeuroEngineering and Rehabilitation, 14(1):82.
Fisher, R. A. (1936). The use of multiple measurements in taxonomic problems. Annals of Eugenics, 7(2):179–188.
Hudgins, B., Parker, P., and Scott, R. N. (1993). A new strategy for multifunction myoelectric control. IEEE Transactions on Biomedical Engineering, 40(1):82–94.
Jolliffe, I. T. (2002). Principal Component Analysis. Springer Series in Statistics. Springer, New York, 2 edition.
McInnes, L., Healy, J., and Melville, J. (2018). Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426.
Phinyomark, A., Khushaba, R. N., Ibáñez-Marcelo, E., Patania, A., Scheme, E., and Petri, G. (2018). Feature extraction and selection for myoelectric control based on wearable emg sensors. Sensors, 18(5):1615.
Rabiner, L. R. and Juang, B.-H. (1993). Fundamentals of Speech Recognition. Prentice Hall, Englewood Cliffs, New Jersey.
Schultz, T. and Wand, M. (2010). Modeling coarticulation in emg-based continuous speech recognition. Speech Communication, 52(4):341–353.
Spacone, G., Frey, S., Pollo, G., Burrello, A., Pagliari, D. J., Kartsch, V., Cossettini, A., and Benini, L. (2026). Silentwear: an ultra-low power wearable system for emg-based silent speech recognition. arXiv preprint arXiv:2603.02847.
van der Maaten, L. and Hinton, G. (2008). Visualizing data using t-sne. Journal of Machine Learning Research, 9:2579–2605.
