Digging into sentence intelligibility: Interactions with talker

Abstract

An implicit feature found in studies of intelligibility and perception is the assumption that noise masks all spoken-sentence stimuli equivalently regardless of the stimulus-talker. For example, when using controlled and normed sentences, such as the IEEE Harvard set, the mean intelligibility across talkers should be affected by masker-noise equivalently. In other words, any particular sentence spoken by one talker should be equivalently intelligible to when it is spoken by another talker using the same style and rate. However, previous studies have shown that there is significant intelligibility variation between talkers in controlled stimuli. Therefore, we predict talker-masker interactions in the IEEE stimuli. We investigated the UAW speech intelligibility dataset which contains typed responses from over 900 native listeners of English to over 604 sentences from the UWNU IEEE sentence corpus presented in three signal-to-noise ratios ( + 2, 0, −2 dB). We calculated the Levenshtein Distance for the listeners’ transcriptions of the sentences. We find that different talkers have different intrinsic intelligibility even when reading the same set of sentences. Moreover, we find an interaction with SNR level and individual talker intelligibility. The present results suggest that individual talker effects in speech stimuli may influence the outcomes of perception and intelligibility studies.