Across- and within-consonant errors for isolated syllables in noise

Area of Science:

• Auditory Perception
• Speech Processing
• Psychoacoustics

Background:

• Speech recognition models often use average error rates.
• The articulation index predicts logarithmic decreases in average error with increased signal-to-noise ratio (SNR).

Purpose of the Study:

• Investigate if the log-linear relationship between error and SNR applies to individual consonants and tokens.
• Identify factors contributing to error rate differences across and within consonant categories.

Main Methods:

• Normal-hearing listeners identified consonant-vowel (CV) syllables.
• Stimuli from 16 consonants and 4 vowels by 14 talkers were presented at 6 SNRs.

Main Results:

• The log-linear relationship held across consonants but not at the token level.
• Across- and within-consonant variability explained most listener errors.
• Adjusting for token-level error thresholds explained 62% of response variability.

Conclusions:

• Speech recognition tests must account for token-level variability, not just averages.
• Understanding token-level differences in normal-hearing listeners can inform diagnostic tests for hearing-impaired individuals.