QIF-T0081

Accelerometer speech reconstruction (vibration-to-audio via surface-coupled MEMS)

Modern smartphone accelerometers (ADXL345, BMI270) have sufficient sensitivity to capture speech vibrations transmitted through surfaces. When a phone lies on a table during a conversation, or is held against the ear during a call, speech vibrations propagate through the phone chassis to the MEMS accelerometer. Ba et al. (2020) used deep learning to reconstruct intelligible speech from accelerometer data alone, achieving speaker identification and keyword recognition. Unlike the gyroscope attack (T0080), accelerometers benefit from direct surface coupling — speech vibrations transmitted through desks, tables, or the user's hand provide stronger signal. Like T0080, accelerometer access required no permission on Android until recent API changes. The combination of gyroscope + accelerometer data (sensor fusion) significantly improves reconstruction quality over either sensor alone.