The influence of linguistic variables on the intelligibility of competing speech signals

Dr Michael Johnston (University of Melbourne) and Dr Ken McAnally (Defence Science and Technology Organisation)

The intelligibility of speech in noisy environments has been the focus of a considerable research effort for half a century. However, this research has predominantly been concerned with the manipulation of purely acoustic properties of the masking noise. The effect of information masking has been considered in only the broadest sense, by comparing the effects of speech and non-speech maskers where the latter are matched to the former for psycho-acoustic properties. To investigate the effect of the linguistic properties of masking speech, that is of syntax and semantics, on the intelligibility of a target voice we used a paradigm involving the presentation of two competing synthesised speech signals. Both signals used the same male-voice speech synthesis algorithm, and were presented through headphones in the same stereo location. One signal was a target question with an obvious answer, and the other was a mask. The subjects’ task was to answer the question. Target sentences were tested under four different masking-sentence conditions: one in which the target and mask were semantically related, one in which they were not semantically related, one in which the mask comprised words in a syntactically illegal order, and one in which masking sentences comprised nonwords. Performance was most accurate in the nonword masking condition, followed by the two syntactically legal masking conditions, with the least accurate performance in the syntactically illegal condition. The difference between the syntactically legal and illegal masking conditions suggests that a syntactically legal masking sentence can more easily be treated as a coherent sound object than an illegal sentence, perhaps due to a mechanism that parses syntactically legal sentences into such objects prior to selection. The additional masking in the conditions in which masking sentences were comprised of words, compared to the nonword masking condition, can be attributed to competition for processing resources in the phonological lexicon. In the case of a masking sentence containing words, such competition might occur between words in the target signal and those in the masking signal, with this competition reduced or eliminated when the masking signal comprises nonwords. However, because nonwords are phonemic in structure, they should constitute acceptable input to the word recognition system, presumably producing some activation within the phonological lexicon, which should result in at least some interference. Thus, an alternative explanation for the observed performance difference between these conditions is that the competition occurs at the stage of output from the phonological lexicon.