Nanowire Strain Sensor-Based Word Recognition for Speech-Impaired Users
Eric Zhang, Shuang Wu, Alex Shen, Zahid Syed, Yonggang Zhu, Xipeng Shen · 2024
Millions of people around the world gra p-ple with speech loss due to various medical conditions. This significantly hinders their ability to communicate and express themselves. To address this challenge, our re-search delves into cutting-edge nanomaterial- based high-performance strain sensors designed specifically to detect intricate human facial muscle movements associated with speech. The strain sensor is soft and flexible, which makes it a good candidate for integration into wearable devices. Our experiments reveal that the strain sensors can pick up distinct signal patterns during the articulation of words and sentences. We observed signal patterns that remained consistent both with and without voice. Furthermore, we discerned and interpreted these signals as words with modern time series machine learning algorithms. By testing 21 popular words, our findings revealed promising outcomes, showing recognition accuracy ranging from 47.4% to 80% across 10 speakers using the K-Neighbors Time Series Classifier. In addition, we compared the classification accuracy of five popular time series machine learning algorithms. Three random forest classifiers showed comparable accuracy with the K-Neighbors Time Series Classifier, while the Time Series Support Vector Classifier showed lower performance. These findings pave the way for a future where individuals with speech impairments can regain a powerful mode of communication through this innova ti ve technology.