In this paper we present the use of confidence measures to improve the comprehension of automatic transcription by hard of hearing. The framework consists in live shows or live streams automatically transcribed by a large vocabulary speech recognition system. We have defined local confidence measures that can be estimated as soon as possible without having to wait for the recognition process to be completed. They have achieved results close to a reference post-processed measure computed on the whole signal and known to be the currently best accurate measure. We have then conducted an experiment to test the contribution of our confidence measure to improve the comprehension of an automatic transcription containing errors. We have thus introduced several modalities to highlight words of low confidence in this transcription and we have shown that these modalities can improve the comprehension of automatic transcription.