Rethinking Mean Length of Utterance for Aided Utterances

Abstract

AIM

Mean length of utterance (MLU) has long been used as a developmentally sensitive measure in early spoken language. However, issues such as symbol relevance and word order likely affect the use of this measure when children use picture symbols to communicate, and MLU – as it’s typically applied to aided utterances – does not account for these unique issues. Therefore, we used a large corpus of aided language utterances to examine the development of a new, and hopefully improved, MLU for use with aided utterances.

We asked how these four MLU versions differ in magnitude, distribution, and variability, and to what extent these variables contribute to differences in MLU values for preschoolers who use aided AAC.

METHOD

The data from 24 children ages 3;0 to 4;11 who used AAC were analyzed. The children had a range of disabilities such as cerebral palsy, cleft palate, and childhood apraxia of speech. All participants had poor intelligibility but receptive language skills that were relatively intact. The current study included data collected following 3 and 4 months of intervention to capture emerging proficiency.

We used a within-sample comparative design to evaluate the alternative operationalizations.

The key variables were operationalized as follows: (A) Listed utterances contained at least three consecutive symbols from the same semantic class without evidence of syntactic structure (e.g., DOG CAT HORSE). Children likely produce these types of utterances because on many AAC devices, semantically related words are listed adjacently in rows or columns. We hypothesize that including these ‘listed’ utterances will inflate MLU. (B) For symbol relevance, symbols that were selected unintentionally or were unrelated to the topic were coded as not relevant. Aided systems allow children to select irrelevant symbols at will, a feature of language that is seldom seen in spoken language and also likely inflates MLU. (C) Word order scores were assigned as follows: 0 points for utterances with no discernable word order, 0.5 points for utterances with partial or emerging word order, and 1 point for utterances without word order errors. Word order issues are well-established in the AAC literature.

Interrater data reliability measures ranged from acceptable to excellent for all variables.

RESULTS

Consistent with our expectations, values were highest when all symbols were included and decreased as the calculations accounted for listing, communicative relevance, and word order. These patterns suggest that each step in refining the MLU calculation results in a more constrained estimate of aided language output.

The mean (SD) for each MLU iteration:

(a) total symbols with listing = 3.02 (0.90);

(b) total symbols without listing = 2.98 (0.94);

(c) total relevant symbols = 2.53 (0.92); and

(d) weighted relevant symbols (i.e., W-MLUSym) = 1.95 (0.81).

Thus, with each iteration, the MLU decreased, with relevance and word order scores resulting in the largest decrease in MLU values. The final value – WMLUSym – indicates that not all symbol combinations were fully formed or well-ordered utterances.

CONCLUSION

Across analyses, a consistent pattern emerged: as additional constraints were applied to account for listing, symbol relevance, and word order, MLU values became progressively smaller but more interpretable and more strongly aligned with other indicators of language ability. Taken together, these findings suggest that applying spoken-language MLU methods to AAC tends to capture how many symbols are selected, rather than how effectively they are used to create meaningful, organized utterances. More broadly, these results underscore the need for modality-specific adaptations of language measures, as features inherent to aided communication systems introduce sources of variance that are not present in spoken language and must be accounted for to obtain valid estimates of language development.

REFERENCES

Binger, C., Harrington, N., Hahs-Vaughn, D., & Kent-Walsh, J. (2026). AAC generative language intervention: A randomized controlled trial. Journal of Speech, Language, and Hearing Research, 69(6), 2654-2678. https://doi.org/10.1044/2026_JSLHR-25-00516

Kent-Walsh, J., Harrington, N., Hahs-Vaughn, D., & Binger, C. (2025). Generative Language Intervention for Young Children With Down Syndrome Using Augmentative and Alternative Communication: A Randomized Controlled Trial. Language, Speech, and Hearing Services in Schools, 56(3), 542–564. https://doi.org/10.1044/2025_LSHSS-24-00117