Convolutional Neural Network-Based Prediction of a French Modified Rhyme Test Recorded with a Body-Conduction Microphone

In the process of improving communication systems, objective intelligibility predictors could circumvent time-consuming listening tests. However, some existing methods are not suitable for devices like body-conduction microphones. The recent development of deep learning for audio applications provides promising solutions for intelligibility prediction with such systems. The goal of this paper is the evaluation of a prediction method based on convolutional neural networks. We train it to retrieve the outcome of a French Modified Rhyme Test (MRT) with controlled artificial noise and low-pass filtering. Its performance is analyzed with data that was excluded from training, as well as with recordings from an in-ear microphone. Compared to the Short-Term Objective Intelligibility (STOI) metric, the network’s predictions correlate better with the listening test. Furthermore, we are able to predict the effect of the speaker’s gender and the tested consonant, unlike the STOI.