Predicting the intelligibility of noisy and enhanced binaural speech
Modern hearing aids make use of increasingly sophisticated signal enhancement techniques such as noise reduction, de-reverberation and binaural beamforming. While the availability of such techniques allows for increasing the performance of hearing aids, it also complicates the task of designing them by introducing many additional design decisions, each of which has to be investigated and validated. Such validations are often done through speech intelligibility tests with many human subjects. These are time consuming and expensive to carry out. Such considerations have recently led to an increasing interest in the application of objective speech intelligibility measures in the context of hearing aids. If such measures can be used to predict the advantage of different signal processing techniques, the required number of speech intelligibility tests can be greatly reduced. For a speech intelligibility measure to be applicable for evaluation of hearing aids, it must be able to take account of the factors which influence the intelligibility of a hearing aid user. In particular, this includes the influence of 1) non-linear processing as carried out by the hearing aid and 2) binaural advantage from spatial separation between target and masker (which must be expected to depend on hearing aid processing as well). We propose and evaluate a binaural intelligibility measure for noisy and enhanced speech. The measure is based on combining the short term objective intelligibility (STOI) measure with an equalization cancellation (EC) stage. The STOI measure is a monaural intelligibility measure, which has been shown to predict well the influence of several speech enhancement techniques. The EC stage is a simple model of how the brain obtains a binaural advantage. By combining the EC stage and the monaural STOI measure, we obtain a measure which is theoretically capable of predicting the full effect, on a normal hearing subject, of wearing a set of hearing aids. We evaluate the proposed measure against a range of listening tests and show that the measure can accurately predict 1) the binaural advantage of spatial separation between a frontal target and an additive noise source in the horizontal plane, 2) the impact on binaural advantage obtained by processing the left and right ear signals independently with ideal binary masks (IBMs) and 3) the impact of wearing a hearing aid with beamforming in conditions with multiple interferers.