State of Kids' Speech Practice 2026

Methods

Data source. SpeechLP is a speech practice app in which children say target words aloud and SpeechLP's on-device phoneme model judges whether the target sound was produced. Each practice card (one word) is logged as an analytics event. We analyzed every practice card logged between January 22 and October 7, 2026 (Toronto time). SpeechLP moved scoring to its on-device phoneme model in January 2026, and the window starts after that change. Practice plans, assigned homework and other data stored outside the analytics system were not available for this report.

Exclusions. We removed internal, staff, test, demo and guest accounts using the same rules as SpeechLP's internal reporting, and cards with no app version recorded. We excluded screening and onboarding cards (they are assessments, not practice) and cards that target more than one sound at once. Per-child figures use family (parent) accounts only; accounts set up by speech-language pathologists were excluded from them. After exclusions, the per-child analysis covers about 950 children and about 61,000 practice words.

Definitions. A child is one child profile within one account. Age is the age entered for the child at their first card in the window, grouped as 2 to 5, 6 to 8 and 9 to 12 years; children with no age, or an age outside 2 to 12, are included in overall figures only. A session is a run of practice within one app session, split wherever more than 30 minutes pass between cards. Weeks are counted from each child's first card (week 1 is days 0 to 6). An active week is a week with at least one session. Session length is the time from the start of the first card to the end of the last card. Sounds are written with familiar symbols: /r/ stands for the English r sound (IPA /ɹ/), and "th", "sh" and "ch" for the sounds in think or this, ship and chip.

Model-scored accuracy. A practice word counts as correct when SpeechLP's on-device phoneme model accepted it within that card's attempts. The number of attempts per card is not logged, so model-scored accuracy is not first-try accuracy, and it is not a clinician's judgment. The section How accurate is the scoring model? describes what is known about how the model compares with a speech-language pathologist.

Statistics. We report medians or means across children, with 95% confidence intervals from a bootstrap that resamples whole children (2,000 resamples), so that children who practice a lot do not dominate. Changes in model-scored accuracy compare each child with themselves on the same sound, and are reported in percentage points. Three separate analyses were run on the same frozen data extract with written conventions, and headline figures were re-derived by a separate verification analysis from its own data pull before publication. All of this work was done by or for SpeechLP, using AI-assisted analysis tools; every figure comes from saved, re-runnable code, and the published wording was checked against the data before release. Where these figures differed slightly because of small differences in definitions, we publish the figure at a precision on which both agree (for example, "about 950 children"). Two follow-up analyses on practice amount and frequency were specified in writing before they were computed and are reported regardless of their direction.

Privacy. Only aggregate statistics are published. Every published figure from practice data covers at least 5 children, with secondary suppression so that no smaller group can be worked out by subtraction. No audio, names, words as said, individual timelines or exact dates tied to small groups are published.

Results

The results cover about 950 children and about 61,000 practice words on family accounts between January 22 and October 7, 2026. The first sections describe how much, how consistently, how long and when children practiced, and which sounds they worked on. The last section looks at whether model-scored accuracy changed across a child's sessions on a sound.

How accurate is the scoring model?

Every accuracy figure in this report comes from SpeechLP's on-device phoneme model, which listens to each attempt and decides whether the target sound was produced. A practice word counts as correct if the model accepted it within the card's attempts. It is not first-try accuracy and it is not a clinician's score, so it should be read as a practice signal, not a clinical measure.

The only published paired comparison. In a small 2025 internal comparison, a speech-language pathologist transcribed 49 recorded child word attempts from very few speakers, each containing a target sound such as /s/, /z/, /f/, /v/, "th", "sh" or "ch". We then checked whether each system's output agreed with the clinician on whether the target sound was produced. The phoneme model SpeechLP selected in that comparison, an earlier version than the one in the app today, agreed with the clinician on 82% of the attempts (Cohen's kappa 0.59, 95% CI about 0.34 to 0.81). A general-purpose speech-to-text system, of the kind that scored practice in early versions of the app, agreed 57% of the time (kappa 0.04, no better than chance).

How does that compare? A systematic review of automated speech analysis tools for children applied an 80% agreement threshold, based on commonly accepted agreement between two human raters of 75 to 85% for perceptual speech judgments (McKechnie et al., 2018, citing Charter, 2003, and Cucchiarini, 1996). The point estimate clears that bar, but the sample is small, the lower end of its uncertainty range falls below 80%, and a kappa of 0.59 means agreement well above chance but far from perfect.

This comparison has real weaknesses. It used a small set of recordings from very few speakers, scored by one speech-language pathologist, and when three SLPs transcribed two of the same recordings, they disagreed on some sounds. The scoring rule checked whether the sound appeared anywhere in the word rather than in its target position, and the model tested was an earlier version than the one in the app today. It should not be summarized as "SpeechLP is 82% accurate."

Internal testing. In internal testing, SpeechLP's team, including its speech-language pathologists, checked the model's detection on every target sound across hundreds of trials with about 10 children aged 2 to 7. That testing found detection performing as intended on all target sounds except /s/ produced with a lateral lisp. Detection of lateral /s/ errors is a known weakness under active development. Item-level results from this testing were not recorded in a form we can publish or independently analyze, so no agreement statistic is reported from it. This testing was done by SpeechLP and was not independent.

What comes next. A formal, blinded study of agreement between the model and speech-language pathologists, with a larger and more varied group of children, is planned. Until then, model-scored accuracy is best read as a signal of how practice is going, not as a measure of clinical progress.

What this means for parents

Speech sound difficulties are common: the National Institute on Deafness and Other Communication Disorders estimates that 8 to 9% of young children have a speech sound disorder (NIDCD, 2025). If you are practicing at home, here is what the data suggest.

• Short sessions are common. A typical session in this data lasted a couple of minutes. This report cannot say how much practice is enough, so ask your speech-language pathologist how often to practice and how many words to aim for.
• Keep going past the first week. Most children in this data practiced in a short burst and then stopped. The first week or two is when routines are easiest to lose, so plan the next few sessions ahead.
• Frequency mattered in clinical research. In one clinical trial with preschoolers, children seen three times a week did better than children who had the same number of sessions once a week (Allen, 2013). This report's app data could not confirm a frequency effect, so discuss a schedule with your speech-language pathologist.
• Evenings work for many families. Evening was the most common practice time in this data. Pick a time that fits your routine, such as after dinner or before bath, and keep it predictable.
• /r/ takes time. The r sound is one of the last sounds children learn: in U.S. studies, the age by which 90% of children said it correctly averaged about 5;6 (Crowe & McLeod, 2020). In this data, most children aged 6 and older practiced /r/. If your child is past that age and still working on /r/, a speech-language pathologist can advise you.
• Treat the app's score as a practice signal. SpeechLP's score shows how practice is going; it is not a diagnosis and not a clinician's assessment. Scores move around from session to session, and even a trend over weeks is a model's score, not a clinical measure of progress.
• Keep your speech-language pathologist in the loop. Share which sounds and words your child is practicing and how often. If you have concerns about your child's speech and do not have an SLP, a licensed speech-language pathologist can evaluate your child.

What this means for SLPs

Dose at home is small. Published phonological intervention research typically delivered 2 to 3 sessions a week of about 100 production trials each (Sugden et al., 2018). In this data, the median home session covered 8 practice words and the median child practiced 16 words in an active week. A practice word can include more than one attempt, and attempts are not logged, so production trials are undercounted; even so, the gap between research dose and home dose is large. Reviews of a small evidence base suggest that greater treatment intensity tends to go with better outcomes in speech sound disorders (Kaipa & Peterson, 2016; Williams, 2012), and in one trial with preschoolers, more frequent sessions did better at equal total sessions (Allen, 2013).

Frequency is the weak point. Children practiced in about a third (34%) of their first four weeks (about 780 children), and most practiced in only one of those weeks. Home practice is rarely described in detail in research: of 176 phonological intervention papers, 61 reported parent involvement or home tasks, mostly with limited detail (Sugden et al., 2016), and surveyed SLPs report that following up on homework completion is inconsistent, with school-based SLPs less likely to follow up (Tambyraja, 2020). App practice logs give you a frequency record to review with families at each visit.

What model-scored accuracy can tell you: whether a child is practicing, on which sounds and words, and the direction of their in-app scores on a sound over weeks. What it cannot tell you: first-try accuracy (attempts are not logged), accuracy as you would score it, accuracy in a target position (agreement was checked on presence of the sound in the word), carryover to connected speech, or generalization outside the app. Detection of lateral /s/ errors is a known weakness. A mastery criterion such as 80% across three consecutive sessions should not be applied to model-scored accuracy as if it were clinician-scored.

On progress, the honest reading is narrow. Model-scored accuracy on a sound stayed about the same across sessions, a small difference that does appear is tied to repeated words rather than new ones, and we found no dose-response. Fewer than 4% of children practiced long enough to test a three-session mastery criterion at all. These data neither show that home practice in SpeechLP changes how a child speaks nor show that it does not; with home doses this small and this short, a null in observational data is expected either way. A blinded, controlled study with SLP-scored untrained probe words is planned.

Keeping families practicing is a practical lever. Only about a third of children practiced in the app again after their first week (n = 855). Building the next two or three sessions into the plan, checking in early, and choosing words with the child's carryover goals in mind are practical steps that do not depend on any app.

Target selection. /r/ dominated practice for children aged 6 and older, at or past the average age (5;6) by which 90% of children in U.S. norm studies produce it (Crowe & McLeod, 2020), while younger children practiced a broad mix of early sounds. Word selection is itself an active ingredient in phonological learning (Storkel, 2018), so the words assigned for home practice deserve the same attention as the sound.

Limitations

This is a descriptive report of product data. Its main limitations:

• Observational data with no control group. Nothing in this report shows that practice in SpeechLP causes change in a child's speech. A blinded, controlled study with SLP-scored untrained probe words is planned.
• No IRB review and no controlled trial. The data were collected in the normal course of using the app and analyzed in aggregate.
• Internal analysis and review. The report was produced and funded by SpeechLP, the analysis was done by SpeechLP, and the clinical reviewers are SpeechLP co-founders and co-CEOs, so the review was not independent. No external peer review has been done.
• Model-scored, not clinician-scored. Accuracy comes from SpeechLP's on-device phoneme model. Its agreement with clinicians has been checked only in a small 2025 comparison of an earlier model version and in internal testing without recorded item-level results; no paired comparison of the model in the app today has been published. Detection of lateral /s/ errors is a known weakness under active development.
• Attempts are not logged. A word counts as correct if it passed within the card's attempts, so model-scored accuracy is not first-try accuracy, and more retries or prompting would raise it.
• Family accounts only. Per-child figures exclude children on accounts set up by speech-language pathologists, so they may not describe children who practice under an SLP's account.
• Start of the window. Children who began using the app before January 22, 2026 are counted from their first card in the window, so for some children the first week in this data is not their true first week.
• Account type is current-state. Whether an account is a parent account, linked to an SLP or an SLP account reflects its setting today, not necessarily at the time of practice.
• Analysis choices. The within-child comparison of model-scored accuracy is sensitive to the minimum number of words a session must contain, a threshold that was not set in advance (see Results). No improvement claim is made.
• The app changed over the window. Scoring, content, games and experiments changed during 2026, which affects comparisons over time.
• Data stored outside the analytics system, such as practice plans and assigned homework, were not available, so we cannot compare practice with what was assigned.
• Self-selected users. Families who choose and keep using a practice app differ from families who do not, and children who kept practicing differ from those who stopped.
• No diagnosis or clinical data. We do not know children's diagnoses, severity, outside therapy, or who helped them practice.
• Practice outside the app is invisible, so weeks without app practice are not necessarily weeks without practice.
• Ages are as entered by families, and age bands with too few children are not reported.

How these numbers relate to earlier SpeechLP posts

This report is the first time SpeechLP has published statistics on how children practice in the app. Figures that appear in SpeechLP's app store listing, website, ads or social posts, such as word and game counts, describe the product rather than practice, cover different periods and accounts, and were not produced with this report's exclusions and methods, so they should not be compared with the numbers here. When citing SpeechLP practice data, cite this report, with its window, its exclusions and the number of children behind each figure.

Data availability

State of Kids' Speech Practice 2026: aggregate tables

Aggregate statistics on at-home speech sound practice in the SpeechLP app from January 22 to October 7, 2026, mainly covering about 950 children on family accounts: practice volume, consistency, session length and timing, sounds practiced by age, and within-child change in model-scored accuracy, plus the 2025 model comparison. Each row is one published statistic with its estimate, 95% confidence interval and number of children; game completion by account type appears only as ranges in the notes. Every cell based on practice data covers at least 5 children. Released under CC BY 4.0.

The tables behind every chart and headline figure are available as a CSV file, and the full report as a PDF, using the download buttons on this page. Both are released under the Creative Commons Attribution 4.0 International license (CC BY 4.0): you may share and adapt them with credit to SpeechLP.

To protect children's privacy, row-level data are not released, and every published cell based on practice data covers at least 5 children (or 5 accounts). Questions about the methods can be sent to support@speechlp.com.

Period covered

January 22, 2026 to October 7, 2026

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Cost

Free to access

Variables in the dataset

Practice words per active week (words per week)

Practice words (one card each) per child in weeks with at least one session; median and mean across children.

Share of first four weeks with practice (percent)

Share of a child's first four weeks with at least one practice session, among children with at least 28 days of follow-up.

Children still practicing (percent)

Share of children with any SpeechLP practice in week 2, 4 or 8 after their first card, or later.

Session length (minutes)

Minutes from the start of the first card to the end of the last card in a session, and practice words per session.

Session start time (percent)

Share of practice sessions starting in each block of the device's local time, including evening (17:00 to 21:59) and weekend shares.

Sounds practiced (percent)

Share of practice words and share of children for each target sound, overall and by age band.

First-round game completion (percent)

Share of game starts whose next game event was a finished game, by account type recorded today. Given only as ranges (more than half, fewer than half) in the notes, not as estimates.

Model-scored accuracy (percent)

Share of practice words that SpeechLP's on-device phoneme model accepted within the card's attempts. Not first-try accuracy and not clinician-scored.

Within-child change in model-scored accuracy (percentage points)

Difference between a child's later sessions and first session on the same sound, with sensitivity analyses.

As heard on

• The ABA Speech Podcast: Episode #281: Carryover For Speech Sound Practice with Shail Silver, hosted by Rose Griffin May 19, 2026
• Teaching Matters Podcast: Speech LP: An AI Tool for Language Instruction, hosted by Scott Titsworth April 7, 2026

References

1. Allen, M. M. (2013). Intervention efficacy and intensity for children with speech sound disorder. Journal of Speech, Language, and Hearing Research, 56(3), 865-877. https://doi.org/10.1044/1092-4388(2012/11-0076)
2. Charter, R. A. (2003). A breakdown of reliability coefficients by test type and reliability method, and the clinical implications of low reliability. The Journal of General Psychology, 130(3), 290-304. https://doi.org/10.1080/00221300309601160
3. Crowe, K., & McLeod, S. (2020). Children's English consonant acquisition in the United States: A review. American Journal of Speech-Language Pathology, 29(4), 2155-2169. https://doi.org/10.1044/2020_AJSLP-19-00168
4. Cucchiarini, C. (1996). Assessing transcription agreement: Methodological aspects. Clinical Linguistics & Phonetics, 10(2), 131-155. https://doi.org/10.3109/02699209608985167
5. Kaipa, R., & Peterson, A. M. (2016). A systematic review of treatment intensity in speech disorders. International Journal of Speech-Language Pathology, 18(6), 507-520. https://doi.org/10.3109/17549507.2015.1126640
6. McKechnie, J., Ahmed, B., Gutierrez-Osuna, R., Monroe, P., McCabe, P., & Ballard, K. J. (2018). Automated speech analysis tools for children's speech production: A systematic literature review. International Journal of Speech-Language Pathology, 20(6), 583-598. https://doi.org/10.1080/17549507.2018.1477991
7. National Institute on Deafness and Other Communication Disorders. (2025, July 8). Quick statistics about voice, speech, language. https://www.nidcd.nih.gov/health/statistics/quick-statistics-voice-speech-language
8. Storkel, H. L. (2018). Implementing evidence-based practice: Selecting treatment words to boost phonological learning. Language, Speech, and Hearing Services in Schools, 49(3), 482-496. https://doi.org/10.1044/2017_LSHSS-17-0080
9. Sugden, E., Baker, E., Munro, N., & Williams, A. L. (2016). Involvement of parents in intervention for childhood speech sound disorders: A review of the evidence. International Journal of Language & Communication Disorders, 51(6), 597-625. https://doi.org/10.1111/1460-6984.12247
10. Sugden, E., Baker, E., Munro, N., Williams, A. L., & Trivette, C. M. (2018). Service delivery and intervention intensity for phonology-based speech sound disorders. International Journal of Language & Communication Disorders, 53(4), 718-734. https://doi.org/10.1111/1460-6984.12399
11. Tambyraja, S. R. (2020). Facilitating parental involvement in speech therapy for children with speech sound disorders: A survey of speech-language pathologists' practices, perspectives, and strategies. American Journal of Speech-Language Pathology, 29(4), 1987-1996. https://doi.org/10.1044/2020_AJSLP-19-00071
12. Williams, A. L. (2012). Intensity in phonological intervention: Is there a prescribed amount? International Journal of Speech-Language Pathology, 14(5), 456-461. https://doi.org/10.3109/17549507.2012.688866