Interpretable Machine Learning-Enabled Time-Resolved Speech Command Recognition Using Triboelectric Laminate Acoustic Sensors

The growth of wearable electronics has intensified the search for sustainable, high-performance materials for integration into smart sensing platforms. Conventional piezoelectric polymers, particularly fluorinated systems, offer excellent electromechanical transduction; however, they pose environmental and end-of-life challenges. Here, we demonstrate a recycled expanded polystyrene (PS) triboelectric laminate as an acoustic neuromorphic sensor. The multilayered structure of the triboelectric laminate enables a mechanoelectric response similar to that of piezoelectric polymers. The triboelectric laminate sensor was integrated into the face mask and generated current signals in the range of 4 nA (peak-to-peak) with distinct features in response to speech signals of various words. To exploit the recorded signals for dynamic speech recognition, we develop an interpretable physics-guided staged ML framework named Triboelectric Lexical Network (TriboLexNet), which learns the spectral signatures and combines time-frequency descriptors embedded into a random-forest framework. TriboLexNet is capable to recognize the specific word in the sentences, with a held-out segment-level accuracy of 99.4 ± 0.9% and a per-recording keyword-detection accuracy of 89.6 ± 6.4% under strict recording-level data partitioning, as well as locating it in terms of an interpretable probability distribution. Our approach establishes a viable platform for sustainable triboelectric acoustic neuromorphic sensing and demonstrates how physics-guided ML can be utilized in dynamic speech recognition with limited data.