Inside Speech AI: four AI layers, two interfaces, one loop of care
What actually happens inside Speech AI? A plain-language tour of the four technical layers and the two connected interfaces that keep clinicians and families on the same page.
Speech AI is often described in one sentence, AI-assisted screening and neurodevelopmental care, but under the hood it is four distinct technical layers, joined by two interfaces that keep specialists and families connected. Here is the plain-language tour.
Layer 1: speech & audio models
Acoustic and phonetic models analyse consented Bengali child speech (vocalisation, pronunciation, word use, response delay, fluency and language patterns) to flag communication-disorder risk. This screening core is the heart of the platform, and the first thing validated against specialist judgment.
Layer 2: interaction analysis
Fine-tuned computer-vision models read facial attention, eye contact, gesture, participation and engagement from short, guided caregiver-child interaction videos. It is developed and tested as a staged research component: structured observation, not surveillance.
Layer 3: growth-stage mapping
A developmental engine compares observed communication, social, cognitive, behavioural and functional indicators against age-appropriate milestones, producing a clear growth-stage profile that anchors every plan.
Layer 4: the therapy engine
A rules-plus-machine-learning engine generates individualized therapy modules and home-practice activities from the growth-stage profile. Therapists review, adjust and confirm every plan before it reaches a family.
Two interfaces, one record
- The clinician dashboard runs AI scoring, plan generation, validation and progress oversight.
- The Bengali guardian app delivers screening prompts, home-practice activities and progress capture.
- Data flows both ways, so the centre keeps oversight of what happens at home between sessions.
Throughout, clinicians and guardians remain in the loop, reviewing and confirming outputs.
— The Speech AI design principle
The loop is the point: screening shapes the growth-stage map, the map shapes the plan, and tracked progress refines the plan again, with a specialist confirming every important step.
