BasaBata: 60-second oral reading miscue and Phil-IRI screener for Grade 1 to 3 learners

Problem Statement

World Bank and EDCOM II assessments show that up to 90 percent of 10-year-old Filipino children experience learning poverty—unable to read and understand a simple age-appropriate story. To diagnose struggling readers early, the Department of Education mandates the Philippine Informal Reading Inventory (Phil-IRI) across all public elementary schools. However, a Grade 1 to 3 public school teacher with 45 to 60 pupils must sit with each child one-on-one, hold a stopwatch, and manually mark 7 types of oral reading miscues (mispronunciation, substitution, omission, insertion, repetition, transposition, and reversal) on a paper score sheet while managing the rest of the class. Administering and computing Phil-IRI takes weeks of instructional hours, leaving teachers exhausted and little time for actual reading remediation.

Proposed Solution

BasaBata turns a 3-week manual Phil-IRI assessment into a 60-second smartphone reading check for public school teachers, barangay reading volunteers, and parents:

1. 60-Second Oral Reading Capture: The pupil reads a graded 50-to-80-word Tagalog, Cebuano, or English Phil-IRI passage displayed on the screen (or from a printed card) into the teacher’s phone microphone.
2. Phoneme-Level Forced Alignment & Miscue Detection: Using Connectionist Temporal Classification (CTC) forced alignment over Wav2Vec 2.0 / Whisper phoneme representations, BasaBata aligns the child’s spoken syllables (pantig) against the reference passage text at 20-millisecond resolution. It automatically marks every Mispronunciation, Omission, Substitution, Insertion, Repetition, and 3-second Hesitation, and computes exact Words Correct Per Minute (WCPM) and the official DepEd Phil-IRI Word Recognition Level (Independent: 97–100%, Instructional: 90–96%, Frustration: $\le 89%$).
3. Instant Marungko Approach Remediation Card: Instead of just labeling a child “Frustration Level,” BasaBata pinpoints the exact consonant-vowel clusters or diphthongs the child stumbled on (for example, ng-, -ay, pl-, or b/d reversals) and generates a 5-minute Marungko Approach phonics drill sheet and SMS parent homework card for that specific child.

Target Market

• Primary users: Over 500,000 public elementary school teachers (Grades 1 to 6), DepEd National Reading Program / Catch-Up Fridays coordinators, and parents of early readers.
• Public-sector and NGO buyers: DepEd curriculum and division offices, Local School Boards (using the municipal Special Education Fund under RA 7160), Synergeia Foundation, USAID ABC+ / early literacy coalitions, and Teach for the Philippines.
• Community benefit: Gives every public school teacher back 30+ hours of manual testing per quarter while catching reading difficulties and dyslexia red flags in Grade 1 before a child falls behind permanently.

How AI Is Used

• CTC phoneme-level forced alignment (Wav2Vec 2.0 / Whisper): Standard speech-to-text models auto-correct a child’s reading mistakes using language-model priors—which ruins reading assessment. BasaBata uses acoustic phoneme alignment without autocorrect so it hears exactly what the child pronounced ("ba-ta" vs "da-ta" or skipped syllables).
• Automated DepEd Phil-IRI formula engine: Applies the exact Phil-IRI Manual 2018 quantitative and qualitative rubrics, combining Word Recognition score with 5 spoken comprehension questions to output the official learner profile.
• Targeted phonics drill generation: Maps the child’s specific phoneme confusion matrix into age-appropriate Tagalog (Marungko sequence: m, s, a, i, o, b, e, u…) or English decodable micro-stories.

Community & Ecosystem Integration

• 100% aligned with DepEd Phil-IRI & MATATAG National Reading Program: Produces the exact Individual Summary Record (ISR) and Class Reading Profile tables required by DepEd division offices.
• Supports Catch-Up Fridays: Groups the class into 4 targeted phonics clusters automatically so the teacher knows which 8 pupils need Marungko syllable blending vs. comprehension coaching on Friday.
• Parent SMS reading loop: Sends a free Tagalog SMS to the parent’s phone with 5 practice words the child struggled with that day.

Demo Scenario (3 minutes for judges)

1. Display a Grade 2 Tagalog Phil-IRI passage (“Ang Munting Gamugamo”, 62 words) on the screen.
2. A volunteer reads the passage aloud for 45 seconds, intentionally making 4 common early-reader miscues: skipping a syllable ("gamugamo" $\rightarrow$ "gamo"), substituting "b" for "d", repeating a word twice, and pausing for 4 seconds on "liwanag".
3. Within 3 seconds of finishing, BasaBata displays the color-coded Phil-IRI Running Record: every miscue is marked with standard DepEd proofreader symbols above the word, speed is scored at 48 WCPM, and Word Recognition is calculated at 91.9% (Instructional Level).
4. Click Generate Remediation: BasaBata identifies the b/d reversal and multi-syllable skipping pattern, prints a 5-minute Marungko practice card, and updates the teacher’s Class Reading Profile table automatically.

30-Hour Build Plan

Suggested stack: Python FastAPI, PyTorch (torchaudio MMS / Wav2Vec2 CTC forced alignment + Whisper timestamped word alignment), Gemini API for Marungko drill generation, Next.js PWA, PostgreSQL. Public data / APIs: DepEd Phil-IRI Manual (2018) public passages and scoring rubrics, Meta MMS (Massively Multilingual Speech) Tagalog/Cebuano acoustic models, Common Voice Tagalog dataset.

Research & Literature References

1. Lingyun Gao et al., “Reading Miscue Detection in Primary School through Automatic Speech Recognition,” Proceedings of Interspeech 2024, pp. 1180-1184, 2024. DOI: 10.21437/Interspeech.2024-1180
   • How BasaBata uses this study: Implements the paper’s self-supervised Wav2Vec 2.0 and Whisper phoneme-level forced alignment pipeline to detect primary-school oral reading miscues (substitutions, deletions, insertions, and hesitations) without language-model auto-correction masking the child’s errors.

Risks & Privacy Notes

• Child voice data protection (RA 10173): Audio recordings of minors are processed in memory to generate the Phil-IRI running record and phoneme confusion counts, then deleted immediately. No raw child voice files are stored on external servers or used to train commercial models.
• Screening aid, not clinical diagnosis: Flags for frequent b/d or p/q reversals are presented to the teacher as targeted phonics practice recommendations, never as a clinical medical diagnosis of dyslexia.
• Classroom noise resilience: Uses close-talk headset or phone-handset mic gain normalization to filter background classroom chatter.

AI Use Disclosure (draft)

BasaBata uses Wav2Vec 2.0 / MMS CTC phoneme forced alignment and Whisper word-level timestamps to detect oral reading miscues, and an LLM grounded in the DepEd Marungko phonics sequence to generate personalized remediation cards. AI coding assistants helped build the running-record annotation UI and PDF exporter. All reading level formulas strictly follow the official DepEd Phil-IRI manual.

Rubric Scoring