Hidden Markov Models and Multiple Transfer Functions for Voice Subtraction in an Acoustic Dosimeter

The exposure of an individual to noise is a critical topic in relation to noise-induced hearing loss and noise annoyance. Unfortunately, this exposure is difficult to quantify, because the individual's voice affects the noise itself. A new dosimeter is presented, relying on an accelerometer and a microphone system. The accelerometer (placed behind the user's ear, in close contact with the skull) is moderately sensitive to environmental noise. It can be used as an estimate of the clean voice, which can then be removed from the overall acoustic signal. Eventually, the voice-free signal can be processed to measure the noise exposure in a straightforward manner. In order to obtain an estimate of the clean voice, the accelerometric signal may be mapped to a virtual microphone signal. We apply autoregressive exogenous (ARX) models to realize such a mapping. As the ARX is a time-invariant model, we use hidden Markov models to split the input sequence into segments on a phonetic-class basis. Specialized ARX models are then estimated for each one of the classes. Simulation results and a real-world binaural digital signal processing (DSP) implementation of the system validate the approach to a significant extent.