A Study on Korean Silent Speech Word Recognition Based on Surface Electromyography

Conventional speech recognition technology suffers from degraded performance in noisy environments and has limited applicability for users who face physical constraints that make speech difficult [1, 2]. As an alternative, silent speech recognition technology, which enables communication using facial and jaw muscle signals without actual vocalization, has attracted attention. Surface electromyography (sEMG)-based silent speech recognition infers the intended utterance from the electrical signals of articulatory muscles. Because it does not rely on acoustic signals or external voice transmission, it offers potential privacy and security advantages and can be used as an assistive technology. However, existing studies have mainly focused on English and Chinese, and research on Korean silent speech recognition remains limited [3-5]. Accordingly, this study constructs a single-participant sEMG dataset based on Korean words and compares statistical and deep-learning-based models to evaluate the feasibility of Korean silent speech recognition and analyze performance differences across methodologies.

Five facial and jaw muscle regions associated with articulatory movements were selected: the Zygomaticus muscle group, the Levator/Depressor labii muscle group, the Orbicularis oris, the Masseter, and the Mentalis. These regions have been commonly utilized in previous sEMG-based speech recognition studies and provide signals related to facial and jaw movements. The dataset was collected from a single participant. Ten Korean words were used: chamyeo, dari, gaeul, haneul, hyogwa, jari, maeum, meori, nara, and wonin. Each word was uttered silently 40 times, resulting in 400 sequences. sEMG signals were recorded from five channels, with one channel assigned to each muscle region described above.

The collected signals were processed with a high-pass filter and a 60 Hz notch filter to reduce low-frequency baseline drift and power-line noise, followed by normalization to reduce inter-trial signal variation. The signals were then segmented into frames using a 50 ms window with a 25 ms hop size. Four time-domain features, Mean Absolute Value (MAV), Root Mean Square (RMS), Zero Crossing (ZC), and Waveform Length (WL), were extracted from each frame, yielding a 20-dimensional feature vector (5 channels × 4 features) at each time point.

For model training and performance evaluation, the 40 utterances of each word were randomly divided at a ratio of 7:1.5:1.5, yielding 28 training samples, 6 validation samples, and 6 test samples per word. The evaluation was performed using the held-out test set, which consisted of 60 samples in total, with 6 samples per word. Thus, each test sample corresponded to approximately 1.67 percentage points of accuracy. To compare Korean silent speech recognition performance using sEMG, two statistical models, Quadratic Discriminant Analysis (QDA) and Hidden Markov Models (HMM), and one deep-learning-based model, a Bidirectional Gated Recurrent Unit (BiGRU), were applied.

The models achieved accuracies of 50.00% (30/60) for QDA, 48.33% (29/60) for HMM, and 86.67% (52/60) for BiGRU. Among the three models, BiGRU achieved the highest recognition accuracy, outperforming QDA by 36.67 percentage points. This higher performance may be associated with the ability of the bidirectional architecture to capture temporal information from both forward and backward directions. QDA achieved moderate performance, whereas HMM showed the lowest accuracy. The difference between QDA and HMM was only one correctly classified test sample. The relatively low HMM performance may be related to the limited sequence length, which may have made it difficult to reliably estimate state-transition patterns for individual word classes and may have resulted in predictions biased toward specific classes. The confusion matrices were further examined to analyze the class-wise recognition characteristics of each model. As shown in Figure 1, the BiGRU model exhibited clearly pronounced diagonal components for most words, indicating stable classification performance for most word classes. In particular, chamyeo, dari, haneul, nara, and wonin were classified perfectly (6/6). Most errors were concentrated on gaeul and meori. Gaeul (recall 0.50) was misrecognized as dari in two cases and as meori in one case, while meori (recall 0.67) was confused with jari and wonin. Meori also showed the lowest precision (0.57), as samples from gaeul, hyogwa, and maeum were incorrectly predicted as meori. These confusions may be associated with similarities in the overall intensity and temporal variation patterns of the sEMG signals. As shown in Figure 2, the QDA model showed relatively high recognition performance for chamyeo, maeum, and wonin, with 5 of 6 samples correctly classified for each word. However, nara was misrecognized as jari in five cases and as meori in one case, while haneul was misrecognized as jari in two cases and as maeum in four cases, resulting in 0% recall for both words. Predictions were also concentrated on jari, which accounted for 15 of the 60 predictions. This suggests that similarities in the statistical distributions of the extracted features among certain word classes may have resulted in overlapping decision boundaries, thereby degrading classification performance. Finally, HMM showed the lowest accuracy among the three models. As shown in Figure 3, dari and wonin were classified perfectly (6/6), while chamyeo and maeum achieved 5/6 correct classifications. In contrast, haneul and nara resulted in 0% accuracy. Predictions were biased toward specific words, with maeum accounting for 14 of the 60 predictions and wonin accounting for 13. This pattern may be related to the limited sequence length, which may have made it difficult to reliably estimate time-series state-transition patterns for individual word classes.

Overall, BiGRU achieved the highest accuracy (86.67%), suggesting that deep temporal modeling can capture the temporal characteristics of sEMG signals. QDA, which has a relatively simple statistical structure, performed well for certain words but reached only about half the overall accuracy (50.00%), and HMM showed a similarly low accuracy (48.33%), which may be related to the limited sequence length and the distribution of the data. The major misrecognitions across the classification methods were concentrated among certain word classes, suggesting that similarities in sEMG signal characteristics may contribute to classification errors. Accordingly, future research should consider comprehensive approaches, including expanding the dataset for frequently confused word pairs, adding muscle channels, and applying data augmentation techniques to improve the robustness of temporal information. In addition, because the data were collected from a single participant and the test set contained only six samples per word, generalization to other speakers and the statistical reliability of small accuracy differences could not be established. Future studies should therefore include multiple participants and repeated evaluations to assess speaker-independent performance.

This research was supported by the Gyeongbuk ANCHOR system through the Gyeongbuk ANCHOR CENTER, funded by the Ministry of Education (MOE) and the Gyeongsangbuk-do, Republic of Korea. (2026-ANCHOR-15-107)