Decoding Spanish Vowels from Surface Electromyography Using Dimensionality Reduction

Silent Speech Interfaces aim to decode speech from pre-acoustic physiological signals, providing an alternative to conventional acoustic-based speech recognition. A fundamental question is whether discrete phonetic identities can be recovered from the continuous muscular dynamics underlying speech production. Here, we investigated the feasibility of classifying the five Spanish vowels using surface electromyography (sEMG). We designed and implemented a custom system to synchronously record the activity of three articulatory muscles (Orbicularis Oris, Depressor Anguli Oris, and Mylohyoid) in a native Rioplatense Spanish speaker. Using an incremental analysis approach (thresholding, PCA, and supervised UMAP), we evaluated the separability of the phonetic space. Results showed that the system isolates topological phonetic macro-groups (A, E-I, and O-U). Specifically, extreme motor gestures (A, I, U) were classified with high predictive accuracies (82–100%) under non-linear dimensionality reduction, an algorithmic separation aligning with the phonetic extremes of the traditional vowel triangle. However, intermediate vowels (E, O) showed persistent overlap, suggesting that additional muscular information may be required to fully resolve the vowel space. Overall, these preliminary results support the feasibility of extracting discrete phonetic identities from continuous muscular dynamics.