Our 12-month pilot with NeuroSparks and the University of Dhaka
Speech AI is being validated the hard way: a four-phase, 12-month research pilot inside real therapy centres, supervised by the country's leading academic centre for communication disorders.
A screening tool for children should not be taken on faith. That is why Speech AI is being developed and validated through a structured 12-month research pilot, with clinical partner NeuroSparks and academic supervision from Dr. Hakim Arif, Professor and Founder Chair of the Department of Communication Disorders at the University of Dhaka.
Who the partners are
NeuroSparks is a Bangladesh-based neurodevelopmental and therapy provider offering speech and language therapy, occupational therapy, ABA therapy, special schooling and neuro-medicine support, operating service points in Jatrabari, Narayanganj and Feni. Its therapists face the exact gap Speech AI targets: high screening demand, manual assessment and weak progress tracking.
The Department of Communication Disorders at the University of Dhaka is the country’s leading academic centre for speech-language pathology. It provides the clinical screening criteria, validation methodology and Bengali child-language expertise behind the platform.
The four phases
- Months 1–2, Foundation: team, architecture, security, and the ethics and guardian-consent framework.
- Months 3–6, Core development: Bengali screening models, therapy-plan generator and progress tracking; a working prototype by month 6.
- Months 7–9, Integration & validation: the full platform assembled, with AI outputs validated against specialist assessment.
- Months 10–12, Pilot & reporting: a supervised pilot at NeuroSparks sites, measuring accuracy, time savings, plan acceptance and usability.
≥30%
targeted reduction in screening time, a committed validation goal
What success looks like
Screening accuracy against specialist judgment, sensitivity and specificity targets, AI-clinician agreement, therapy-plan acceptance, guardian usability, and measurable child progress across sessions. If the pilot validates, the model extends district by district, and the flagship research output, a published benchmark of AI screening accuracy for Bengali-speaking children, becomes a first-of-its-kind contribution for a low-resource language.
