Estimation of intelligibility from received arbitrary speech signals with support vector machine

Intelligibility, a vital concern of a speech transmission channel, is quantified using speech transmission index (STI). The standard STI method relies on noisy test signals and thus hinders in-use measurements. Alternative methods to accurately estimate the STI from naturally occurring speech signals have been developed over the past few years using artificial neural networks. This paper presents a new machine learning based method to more accurately estimate the STI from arbitrary running speech using a purpose design signal pre-processor and support vector machines. When compared with the neural network approaches to the problem, the new method exhibits improved estimation accuracy and generalisation capability to arbitrary speech, providing a more applicable method to facilitate in-situ measurements.