Toscano, J.C., Trevino, A., & Allen, J.B. (2013, October). Assessing the assessment: Measuring listeners’ speech recognition errors for specific consonants and tokens. Poster presented at the 2013 Aging and Speech Communication Conference, Indiana University, Bloomington, IN.

Abstract: A critical issue for understanding speech perception, as well as assessing the effects of age-related hearing loss, is identifying the factors that cause listeners to make speech recognition errors. Models of this process, like the articulation index (AI), predict that listeners’ error rates decrease logarithmically with increases in signal-to-noise (SNR) ratio. We were interested in whether (1) this relationship holds for different consonants and individual tokens, and (2) what accounts for differences in error rates at the consonant and token level. We examined listeners’ error rates for syllables (16 consonants and four vowels), produced at six different SNRs, spoken by 14 talkers. We found that the log-linear relationship between error rate and SNR holds for different classes of consonants, but breaks down at the individual token level. Together, these factors account for a major proportion of the variance in listeners’ responses. This demonstrates that normal-hearing listeners are highly affected by the acoustic variability of individual tokens. In addition, the error thresholds for normal-hearing listeners are correlated with those of hearing- impaired listeners, indicating that this threshold provides key information about the recognition of specific sounds for different listeners. Acoustic variability across multiple tokens can cause large differences in listeners’ error rates and consonant confusions. As a consequence, these results suggest that aggregate measures (i.e., those that average across tokens) for assessing speech recognition in hearing-impaired listeners are not sufficient, since they do not take these factors into account.

Download poster: