Ashuza11/ntina-tarl

Code-switched voice agent that assesses children's reading level in their own language and runs level-appropriate practice, based on the Teaching at the Right Level (TaRL) method. Swahili–English and Yoruba–English. Built on Intron's Sahara voice APIs # Ntina A code-switched voice tutor that assesses foundational reading skills in children aged roughly 6 to 8, following the Teaching at the Right Level (TaRL) framework. Swahili-English and Yoruba-English. **Status: partially built.** The reading assessment pipeline (record, score, facilitator confirmation) runs end to end against the real Sahara speech API for both language pairs. The linguistic datasets (text and audio) are published and validated, and a four-provider ASR benchmark is complete. The offline PWA and the drilling/monitoring-record loop beyond a single reading check are not built yet. See Status and roadmap for the specific line between the two. ## Links | Artifact | Status | |---|---| | Text dataset (Hugging Face) | Ash11/ntina-text-dataset-v0 | | Audio dataset (Hugging Face) | Ash11/ntina-voice-dataset-v0 | | Live demo | Not deployed. Runs locally, see Getting started | | Pitch deck | `Ntina — Africomplings Summer School Final Project` | ## The problem More than half of Grade 2 children in resource-limited countries cannot read a single word of a short text. TaRL addresses this by assessing each child individually and grouping them by measured reading level rather than by grade. That assessment is one-on-one, has to be repeated every few weeks, and has to be rebuilt from scratch for every local language it runs in. That rebuilding and repetition is the labor Ntina targets. ## How it works 1. A child reads a known passage aloud on an inexpensive Android phone. 2. Speech-to-text converts the recording into a transcript. 3. Ntina compares the transcript with the known target using fuzzy edit-distance matching. 4. It estimates one of five reading levels: beginner, letter/syllable, word, paragraph, story. 5. A human facilitator confirms the result before it counts. 6. The tutor gives level-appropriate practice prompts and produces a monitoring record. The reading instrument itself is monolingual: a child assessed on Swahili reads Swahili start …