Articu Engine
Speech recognition hears the word. Articu is built to examine the production.
Articu Engine is the speech intelligence layer behind Practice and Clinic—designed around phoneme-level analysis, language-specific rules, uncertainty and clinician review. For clinicians, technical evaluators and future API partners.
Why ordinary ASR is not enough
Understanding a child and evaluating a target sound are different tasks.
General-purpose ASR is rewarded for recovering the intended word. In speech practice, that normalization can erase the exact error a clinician cares about.
| Task | General ASR | Articu Engine |
|---|---|---|
| Intended word | Primary | Context |
| Phoneme production | Secondary / absent | Primary |
| Error type | Usually absent | Structured target |
| Word position | Usually absent | Explicit |
| Language-specific phonology | General | Pack-based |
| Uncertainty | Often opaque | Review state |
| Clinician verification | Not core | Core workflow |
Engine pipeline
From audio to a structured finding.
Every attempt passes through recording-quality checks, alignment, phoneme representation, target comparison, language context and uncertainty estimation before any feedback is shown.
phoneme inventory · phonotactics · dialect rules · developmental context
Structured output
Findings a clinician can read.
Engine output is structured around the target: what was assigned, what was observed, how uncertain the finding is and what should happen next.
Target word demo
“Rabbit” is not enough. How it was produced matters.
Model finding: possible /ɹ/ → /w/ substitution. Low confidence — routed to the clinician review queue instead of an automatic result.
Illustrative example. Released analysis capabilities are documented on the Science page.
Uncertainty is a feature
The safest score is sometimes “needs review.”
Performance varies by sound, age, language, dialect, recording quality and error type. Articu is designed to expose uncertainty rather than compressing every attempt into a confident number.
Abstention
When the system cannot judge, it says so instead of producing a number.
Review routing
Low-confidence attempts become review-queue items with the reason attached.
Clinician verification
Confirmed and corrected labels are always distinguishable from model findings.
Language pack architecture
Context is not optional.
Each language pack carries phoneme inventory, phonotactics, dialect rules and developmental context. The clinician verification loop closes the system.
Language pack
Phoneme inventory · phonotactics · dialect rules · developmental context · target word library · clinical validation status.
Clinician verification loop
Articu proposes → the SLP confirms or corrects → confirmed information enters progress and—only with separate consent—future model improvement.
Known limitations
Published limits, not hidden ones.
Engine performance is evaluated per sound, per language and per age band. Known limitations are documented publicly on the Science page as validation proceeds.
For developers — later
An engine that may become an API.
Articu Engine is designed so its structured findings could be exposed to speech platforms, EdTech and research tools in the future. There is no API product today—validation comes first.
FAQ
Engine questions
Is this different from speech-to-text?
What happens when Articu is uncertain?
Is the JSON example a real API?
Which languages does the Engine support?
Test the engine with real clinical workflows.
Join the clinician pilot and help define which targets and languages the engine must handle first.
Illustrative examples on this page describe target workflows. Released and validated capabilities are documented on the Science page.