EarOE: Enabling Body-Channel Voice Interaction Interface on Earphones via Occlusion Effect

Nowadays, voice input on earphones has become one of the most paramount human-computer interaction approaches. Traditional voice interaction is built on the air channel, which is highly noise-susceptible and suffers from being falsely triggered by nearby competing users. In our work, we design a noise-resistant voice interaction interface on earphones, named <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Ear</small>OE, which takes advantage of the narrow-bandwidth body channel to reconstruct high-fidelity audible speech. Although promising, directly taking the body channel as a voice interaction interface is nontrivial since the limited bandwidth of body-channel speech causes the original timbre information and linguistic content to be lost. To address these issues, we employ an electro-acoustic (EA) model for occlusion effect-based cross-channel correlation analysis. Based on that, we carefully design an attention-based encoder-decoder network to embrace cross-channel correlation for high-quality wide-bandwidth spectrum synthesis. To accommodate individual differences and improve model generalization, we implement a physics-based data augmentation strategy to expand the scale of the training dataset. Through extensive real-world experiments with 28 participants, <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Ear</small>OE can achieve an average Mel-cepstral distance of 8.08, an average modulation spectra distance of 0.82, and an average log-spectral distance of 11.16, outperforming existing solutions.