On Equalisation–Cancellation for Modeling Binaural Intelligibility

Abstract

Measuring speech intelligibility is important for evaluating speech enhancement algorithms, but performing these measurements physically is a time-consuming and costly process. Various metrics have therefore been developed that can predict binaural intelligibility based on clean target and noise (or noisy) input sound signals. One such metric, based on the equalisation cancellation (EC) model, is the binaural speech intelligibility model (BSIM). The way that this metric is formulated math ematically allows a reformulation that is more compact and efficient. In this paper, the EC process is interpreted as an ‘internal’ beamformer, casting the model into a linear algebra framework. This framework makes the BSIM mathematically and computationally more compact and easier to be combined with signal processing or machine learning strategies. Simulations of the original and proposed implementations of the BSIM show that the internal beamformer framework results in similar intelligibility predictions with a typical 5 to 10 times reduction in simulation time. By analysing the modified binaural short-time objective intelligibility (MBSTOI) metric in this framework, it is shown that the presented concepts can be extended to other binaural intelligibility metrics as well.