• The composite aerogel pressure sensors exhibited low hysteresis (13.69%), wide detection range (6.25 Pa-1200 kPa), and cyclic stability to acquire stable and accurate pronunciation signals.
• Over 6888 and 4158 pronunciation signals were collected by the pressure sensor and utilized for training the convolutional neural network model, allowing for accurate recognition of six dialects (96.2% accuracy) and seven words (96.6% accuracy).
• The work emphasizes innovation in both material design and methodology, bridging sensing performance and mechanical properties.
• This research advances silent speech recognition for human–machine interaction and physiological signal monitoring, particularly for non-standard language users.