Skip to content
Open access

Development of an Accelerometer-Based Silent Speech Interface Robust to Head Motion Artifacts Using a Reference-Free Denoising Autoencoder

2026 · IEEE Transactions on Audio, Speech, and Language Processing · Vol 34, pp. 4785-4799 · 0 citations · 59 references

Abstract

Silent speech interfaces (SSIs) employing three-axis accelerometers can provide an effective communication modality for individuals who experience difficulty producing audible speech. However, the accuracy of such systems tends to degrade markedly when users move their heads (e.g., nodding or turning) during silent speech, thereby reducing their reliability in practical scenarios. Apart from constraining users to maintain a fixed head position during silent articulation, no effective solution to this problem has been proposed. To overcome this limitation, we developed a head motion artifact-robust SSI that incorporates a novel denoising autoencoder (DAE) integrating a convolutional neural network with a bidirectional long short-term memory network. To train the DAE, users are required to complete a brief training session consisting solely of head movements without articulation. The key innovation lies in the training strategy: we augmented the training data for the DAE by generating artificial artifacts, enabling the DAE to learn to suppress various motion-induced distortions while preserving genuine articulatory features, so that the classifier—trained on artifact-free data—can operate reliably even when its input contains head-motion artifacts. Evaluations with ten participants over six recording sessions showed substantial performance improvements, with the signal-to-artifact ratio increasing by 1.23 dB (from −1.12 to 0.11 dB) and classification accuracy rising from 56.0% to 82.3%. To our knowledge, the accelerometer-based SSI implemented in this study is the first to demonstrate robust performance under head motion-artifact conditions, highlighting its potential in real-world assistive communication applications.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.