Large-corpus phoneme and word recognition and the generality of lexical context in CVC word perception

Speech recognition may be analyzed in terms of recognition probabilities for perceptual wholes (e.g., words) and parts (e.g., phonemes), where j or the j-factor reveals the number of independent perceptual units required for recognition of the whole (Boothroyd, 1968b; Boothroyd & Nittrouer, 1988; Nittrouer & Boothroyd, 1990). For consonant-vowel-consonant (CVC) nonsense syllables, j ∼ 3 because all 3 phonemes are needed to identify the syllable, but j ∼ 2.5 for real-word CVCs (revealing ∼2.5 independent perceptual units) because higher level contributions such as lexical knowledge enable word recognition even if less than 3 phonemes are accurately received. These findings were almost exclusively determined with the 120-word corpus of the isophonemic word lists (Boothroyd, 1968a; Boothroyd & Nittrouer, 1988), presented one word at a time. It is therefore possible that its generality or applicability may be limited. This study thus determined j by using a much larger and less restricted corpus of real-word CVCs presented in 3-word groups as well as whether j is influenced by test size.

The j-factor for real-word CVCs was derived from the recognition performance of 223 individuals with a broad range of hearing sensitivity by using the Tri-Word Test (Gelfand, 1998), which involves 50 three-word presentations and a corpus of 450 words. The influence of test size was determined from a subsample of 96 participants with separate scores for the first 10, 20, and 25 (and all 50) presentation sets of the full test.

The mean value of j was 2.48 with a 95% confidence interval of 2.44-2.53, which is in good agreement with values obtained with isophonemic word lists, although its value varies among individuals. A significant correlation was found between percent-correct scores and j, but it was small and accounted for only 12.4% of the variance in j for phoneme scores ≥60%. Mean j-factors for the 10-, 20-, 25-, and 50-set test sizes were between 2.49 and 2.53 and were not significantly different from one another.

The j-factor based on a 450-word corpus and tri-word testing confirms and expands on findings from single-word presentations of isophonemic lists and a 120-word corpus. This enhances the generality (external validity) of the notions that j ∼ 2.5 for real-word CVCs, and lexical knowledge enables CVC word recognition based on ∼2.5 independent perceptual units. The robust nature of isophonemic word test outcomes is confirmed by close agreement with those provided by the high-reliability Tri-Word Test. Percent-correct performance was correlated with j but appeared to account for less than 13% of j-factor variance for most scores likely to be encountered in practice. Variability in the size of j suggests individual differences in the ability to take advantage of lexical knowledge in word recognition. The j-factor may be useful to inform rehabilitation needs, intervention content, and outcome assessment, as well as for other clinical applications.