Methodology & Limitations
This page describes how ConductSpeech computes metrics, where normative data comes from, and what the known limitations are. Intended for clinicians and researchers evaluating the tool.
Morpheme Counting (English)
ConductSpeech counts morphemes following standard clinical conventions as described in Brown (1973), the way SALT Software's transcription conventions mark them:
- Inflectional endings (-ed, -ing, plural and third-person -s, possessive -'s, and -en after have) each count as one morpheme, only on a real word, and not on a name (Lucas), a predicate adjective (was tired) or a gerund (went swimming). A name's possessive (Mary's) still counts.
- Contractions count as two morphemes (can't, I'm, he's, let's), except don't, won't and ain't, which count as one.
- SALT slash codes count as typed (dog/s, go/3s). An omission code such as go/*3s is an error and adds no morpheme.
- Words that only look like they have an ending (was, says, Lucas, pants, wedding) count as one morpheme.
- Irregular past tense forms (went, ran, ate) and irregular plurals (mice, leaves) count as one morpheme.
- Catenatives (gonna, wanna, gotta) count as one morpheme.
- Compounds (breakfast, playground) count as one morpheme.
- Diminutives (doggie, birdie) count as one morpheme.
- Fillers (um, uh, er, ah, eh, hm and mm in any spelling, a word coded [FP], and “you know” when it is not the verb) and maze words in parentheses are excluded from word and morpheme counts. Yes and no words such as mhm and uh-huh are words.
- Unintelligible marks (x, xx, xxx, or any run of x) are not words.
Morpheme segmentation uses rule-based word and suffix handling. It is not a probabilistic model. An ending counts only when what is left is a word in a common English word list (SCOWL).
Spanish Samples
A Spanish or Spanish and English sample shows word counts only: utterances counted, words, MLU in words and different words, by the same rules as English. MLU in morphemes, grammatical utterances and the code-switching percent are not shown until each one is checked against a hand-counted set of Spanish samples.
Limitation: English fillers are left out of the word counts, but a Spanish filler such as eh or este counts as a word unless it is in parentheses. No dialect-specific claims are made.
Core Metrics
Advanced Metrics
DSS (Lee, 1974) and IPSyn (Scarborough, 1990) are computed as lightweight rule-based approximations (“DSS-lite” and “IPSyn-lite”). They capture the major scoring categories but may differ from hand-scored reference implementations.
Limitation: For research-grade DSS and IPSyn accuracy, we recommend validating ConductSpeech output against hand-coded samples from your population. These metrics are best used for screening and progress monitoring, not as sole diagnostic criteria.
Age Reference for MLU
MLU in morphemes is compared with one published reference: the control children in Rice et al. (2010), Table 4, conversation samples during toy play in 6-month bands from 2;6 to 8;11. Each row was checked against the paper on November 6, 2026.
- The range on screen is the band's mean plus or minus 1 SD. The paper describes an MLU 1 SD or more below the age mean as a common clinical marker.
- Only conversation and play samples are compared, the kind of sample the reference children gave.
- Grade norms and a Spanish age comparison are not offered: the data available were too few children, or came from TalkBank, whose license bars commercial products.
Z-score estimates come from the same mean and SD. The reference is a population-level comparison and should be interpreted in context: it does not account for dialect variation, bilingual development patterns, or individual circumstances.
Story Words
The Narrative tab counts story words with one of three word lists: Narrative, Expository or Persuasive. Each part of a list, such as setting or initiating event, counts up to 5 different words found in the child's lines. The total is a count of story words with no maximum, not a rubric score. Counts come from keyword rules today. They are a starting point for your own scoring.
Limitation: A word list cannot judge how well a story is told. The count reads English words only, uses the same lists whichever story was told, and may miss story parts told in other words or in a dialect.
Grammatical Error Detection
Errors, and PGU, come from your SALT codes. Nine automatic checks suggest codes in the editor: third-person -s left off or added, subject and verb agreement, overgeneralized past tense (goed, throwed), a missing is, a missing helping verb, pronoun case, a missing a or the, a missing plural -s, and a missing possessive 's. Each check reads a short word list and finds only some errors. A suggestion counts only when you accept it.
Limitation: The checks miss errors outside their word lists, and they may suggest dialectal features (e.g., African American English, Southern American English) as errors. Accept a suggestion only when it is an error in the context of the child's linguistic background. Sensitivity and specificity have not been formally evaluated against a gold-standard reference set.
Report Drafts
The report draft states the sample's measures in sentences. Goal ideas are a fixed starting list for you to edit. Raw audio and patient identifiers are not needed for a draft.
Limitation: Report drafts are intended to save clinician time, not replace clinical judgment. All reports should be reviewed and edited by a qualified SLP before inclusion in official documentation (IEPs, medical records, evaluation reports).
Overall Limitations
- ConductSpeech has not been independently validated in a peer-reviewed study. Clinical validation with gold-standard comparisons is planned.
- Transcription accuracy depends on audio quality. Background noise, overlapping speakers, and low microphone volume degrade results.
- Speaker labels (child or examiner) come from each line's wording, not from the voices, and are often wrong. Check them in the editor.
- The tool is designed for English and Spanish. Other languages are not currently supported.
- Normative data is not available for all age ranges or all populations.
- ConductSpeech is a clinical support tool, not a diagnostic instrument. It should be used alongside professional clinical judgment.
References
- Brown, R. (1973). A First Language: The Early Stages. Harvard University Press.
- Rice, M. L., Smolik, F., Perpich, D., Thompson, T., Rytting, N., & Blossom, M. (2010). Mean length of utterance levels in 6-month intervals for children 3 to 9 years with and without language impairments. Journal of Speech, Language, and Hearing Research, 53(2), 333-349. https://doi.org/10.1044/1092-4388(2009/08-0183)
- Leadholm, B. J., & Miller, J. F. (1992). Language Sample Analysis: The Wisconsin Guide. Wisconsin Department of Public Instruction.
- Lee, L. L. (1974). Developmental Sentence Analysis. Northwestern University Press.
- Scarborough, H. S. (1990). Index of Productive Syntax. Applied Psycholinguistics, 11(1), 1–22.