Paralinguistic Privacy Protection at the Edge

Voice user interfaces and digital assistants are rapidly entering our lives becoming singular touch points spanning devices. These always-on services capture transmit audio data to powerful cloud for further processing subsequent actions. Our voices raw signals collected through these devices contain a host of sensitive paralinguistic information that is transmitted service providers regardless deliberate or false triggers. As emotional patterns attributes like identity, gender, well-being easily inferred using deep acoustic models, we encounter new generation privacy risks by services. One approach mitigate the risk paralinguistic-based breaches exploit combination cloud-based with privacy-preserving, on-device learning filtering before transmitting voice data. In this article introduce EDGY , configurable, lightweight, disentangled representation framework transforms filters high-dimensional identify at edge prior offloading cloud. We evaluate EDGY’s performance explore optimization techniques, including model quantization knowledge distillation, enable private, accurate, efficient on resource-constrained results show runs in tens milliseconds 0.2% relative improvement “zero-shot” ABX score minimal penalties approximately 5.95% word error rate (WER) linguistic representations from signals, CPU single-core ARM processor without specialized hardware.