Speech Privacy Protection based on Optimal Controlling Estimated Speech Transmission Index in Noisy Reverberant Environments

Abstract

Listen

Translate article

Protecting the privacy of conversations containing confidential and sensitive information in semi-open rooms, such as in banks and hospitals, is essential but challenging because their acoustical characteristics, such as room impulse response (RIR) and background noise, are unknown and prone to change. This study proposes a scheme for protecting the privacy of conversations on the basis of feedback control of an estimated speech transmission index (STI). The STI is an objective index related to listening difficulty and is a function of RIR. Without measuring the RIR of the environment where a supposedly private conversation occurs, an STI-estimation method and one RIR model are utilized. The scheme modifies speech signals in such a way that, for an unintended listener, the signals are as unintelligible as they would be in a room with a low STI. To control the late reverberant parameter of the RIR model, a proportional-integral-derivative controller is used whose controller gains are tuned by using a differential evolution optimizer. Simulations and subjective tests were conducted to evaluate the proposed method. The average error between the actual and target estimated STIs is 0.01. Furthermore, the subjective tests showed that the proposed method can effectively protect privacy under different conditions and is preferable to an open-loop control method.