Cochlear implant (CI) users have great difficulty understanding speech in noise. However, some speakers are found to be more intelligible than others. This paper tests whether a glimpse-based model, that was previously used to successfully explain speaker intelligibility in normal-hearing listeners, can also predict speaker intelligibility for CI users. The model employs a front-end that mimics the effects of energetic masking. This is coupled with a back-end that employs statistical models of CI-processed speech to recognise the unmasked glimpses of the target signal. Listening tests were conducted using signals that simulate the effect of hearing speech mixed with speech-shaped noise through a CI, at signal-to-noise ratios ranging from -4 to 6 dB. The intelligibility of 34 different talkers was measured at each noise level. The model is able to explain the variation in the speaker intelligibilities: the correlation between listener and model intelligibilities varies between 0.87 and 0.91 depending on noise level. This is higher than correlations previously found for normal hearing listeners. Our results have the potential to inform future CI signal processing strategies.
Bibliographic reference. Lin, Lin / Barker, Jon / Brown, Guy J. (2015): "The effect of cochlear implant processing on speaker intelligibility: a perceptual study and computer model", In INTERSPEECH-2015, 1566-1570.