Evidence

How Saylo’s scoring is designed

• Saylo knows which practice word the child is attempting, because the speech therapist chose it.
• It compares what it heard with two versions of that word: the correct one and the expected error (“rabbit” and “wabbit”).
• It judges only the target sound in that word, rather than trying to transcribe everything.
• When the evidence isn’t clear either way, it answers “unsure” and gives no buzz, rather than guessing.

This known-word design follows the approach used by the best-studied research systems. It’s much more constrained, and more reliable, than general speech recognition.

What third-party research shows about automated scoring

• A classifier for R, tested on 48 children with R errors who weren’t in its training data, reached an average F1 score of 0.81 against clinician judgments (Benway et al., 2023).
• Used inside real therapy sessions with 5 participants aged 10 and over, agreement with clinicians was lower, about 0.57 to 0.61, close to how well clinicians agreed with each other (0.66) (Benway & Preston, 2023).
• A systematic review found automated tools typically fell below 80% agreement with clinicians on exactly the words children mispronounced (McKechnie et al., 2018).
• For S-type sounds and lisps, published systems are weaker: a 2026 study of Polish-speaking children reported about 73% precision (arXiv 2606.25181). No validated US English lisp classifier was found.

These are other teams’ systems, mostly tested with children older than Saylo’s target age. They show what’s possible, not how Saylo performs.

What Saylo has tested so far

Only software tests with synthetic, computer-generated voices. They check that the pipeline runs end to end. They aren’t clinical validation: synthetic voices aren’t children, and their errors are cleaner than real ones. Saylo hasn’t been tested with children and has no accuracy results to publish yet.

Why Saylo starts with R

• R has the strongest published research on automated scoring, including a validated classifier design.
• Most children say R by about 5½ (90% by 5 years 7 months), so R errors that continue at 6 to 8 are clinically meaningful. When kids learn each sound.
• TH is often still developing at 6 and is a normal dialect feature for many children, and lisps are much harder to score automatically.

More frequent practice tends to help

• In a trial of 54 preschoolers given the same 24 sessions, three sessions a week led to more progress than one a week (Allen, 2013)).
• A systematic review found higher dose or more frequent treatment generally gave better results for speech sound disorders, while noting the research is limited (Kaipa & Peterson, 2016).
• A 2025 review found research protocols use two to four sessions a week, while real services ran mostly once a week, and that home practice can work when a therapist trains and supervises it (Leafe et al., 2025).

Feedback should fade

• In 12 children with R errors, frequent feedback followed by less frequent feedback beat the reverse order (Preston et al., 2018).
• In a wearable voice study with adults, constant real-time feedback changed behaviour only while the device was on; summary feedback held up better after it came off (Van Stan et al., 2017). That was adults’ voices, not children’s speech sounds, so it’s a pointer rather than proof.

That’s why Saylo isn’t designed as an all-day error alarm. The plan is to buzz mostly for successes and let the therapist thin feedback out.

Carryover is the hard part

Children often manage a sound in therapy long before they use it in conversation. Self-monitoring outside the clinic helped sounds carry over in a 1988 study (Koegel et al.), but a 1992 study didn’t find the same effect (Gray & Shelton), so the evidence is mixed. More on carryover.

What still needs to be tested

• Scoring agreement with several speech therapists, by sound, age, word position and dialect, in children aged 6 to 8.
• How much a necklace microphone, background noise and conversation reduce accuracy compared with seated practice.
• Whether a buzz helps children this young learn, and whether it feels like encouragement or pressure.
• Whether families keep practising, and whether gains last after the necklace comes off.
• Eventually, an independently run trial comparing usual care with and without Saylo. Some children’s errors resolve with time, so a fair comparison group matters.

Until then, Saylo is a practice companion for the sounds a child is working on with their SLP. It doesn’t diagnose or treat speech conditions. Full bibliography.