AccMyrinx: Speech Synthesis with Non-Acoustic Sensor

Abstract

The built-in loudspeakers of mobile devices (e.g., smartphones, smartwatches, and tablets) play significant roles in human-machine interaction, such as playing music, making phone calls, and enabling voice-based interaction. Prior studies have pointed out that it is feasible to eavesdrop on the speaker via motion sensors, but whether it is possible to synthesize speech from non-acoustic signals with sub-Nyquist sampling frequency has not been studied. In this paper, we present an end-to-end model to reconstruct the acoustic waveforms that are playing on the loudspeaker through the vibration captured by the built-in accelerometer. Specifically, we present an end-to-end speech synthesis framework dubbed AccMyrinx to eavesdrop on the speaker using the built-in low-resolution accelerometer of mobile devices. AccMyrinx takes advantage of the coexistence of an accelerometer with the loudspeaker on the same motherboard and compromises the loudspeaker by the solid-borne vibrations captured by the accelerometer. Low-resolution vibration signals are fed to a wavelet-based MelGAN to generate intelligible acoustic waveforms. We conducted extensive experiments on a large-scale dataset created based on audio clips downloaded from Voice of America (VOA). The experimental results show that AccMyrinx is capable of reconstructing intelligible acoustic signals that are playing on the loudspeaker with a smoothed word error rate (SWER) of 42.67%. The quality of synthesized speeches could be severely affected by several factors including gender, speech rate, and volume.