Impact of Speaking Styles on the Accuracy of Predicted Speech Intelligibility

Conversational speech used in research studies is not true conversational speech that individuals use in day-to-day communication. Laboratory-created speech materials are read or memorized and repeated and have different acoustic characteristic compared to true conversational speech. It is of interest to investigate how speaking styles (clear speech, lab conversational speech, and natural conversational speech) impact actual (measured) and predicted speech intelligibility in young adults with normal hearing. Two experiments were conducted in the current study. Speech stimuli were created using the contents of the Story Retelling Procedure (SRP) (Doyle et al., 2000; McNeil et al., 2007) produced by a male talker to create stimuli for each of the three speaking styles. Speech recordings were rated by thirty individuals with normal hearing based on how natural speech sounded. There was a strong, positive correlation between the speech recordings and how listeners perceived the naturalness of speaking styles therefore allowing comparison of materials considered clear speech, lab conversational speech, and natural conversational speech. Experiment 1 was designed to investigate if there was any significant difference among speech intelligibility for clear, laboratory conversational, and natural conversational speech in five listening conditions (quiet, +3, 0, -3, and -6 dB SNR). The dependent variable (DV) was proportion correct of identified keywords (i.e., speech intelligibility). The results showed that speaking styles and listening conditions impact measured speech intelligibility. Specifically, there were significant differences in the speech intelligibility between lab conversational speech and natural conversational speech. Moreover, the clear speech speaking style can be used to improve listening performance in challenging listening conditions. Experiment 2 was designed to investigate if the STMI model can accurately predict speech intelligibility for different speaking styles and to evaluate the overall ability of the STMI model to capture speech intelligibility in multi-talker babble noise conditions. The results demonstrated that the current version of the STMI may not be sensitive enough to predict speech intelligibility for different speaking styles when embedded into multi-talker babble.