On the role of discriminative intelligibility model for speech intelligibility enhancement

This paper uses listening tests to directly evaluate two speech pre-enhancement algorithms that were published in earlier work. The models were previously evaluated using purely objective measures. The methods under study aim to increase speech intelligibility by applying an adaptive spectral shaping filter to the speech signal. In both algorithms the shape of the filter is adapted so as to maximise the intelligibility predicted by an objective intelligibility model. The first algorithm uses ‘glimpse proportion’ as the measure of intelligibility (i.e. assuming intelligibility is proportional to the extent of energetic masking). In contrast, the second optimises the score of a statistical ‘microscopic’ intelligibility model that measures the degree of discrimination between the correct interpretation and competing incorrect interpretations of the utterance. Results show that a significant intelligibility gain is obtained when the simple energetic masking model is employed, whereas the discriminative model currently fails to provide any intelligibility improvement.