Methodology

The phone records and uploads audio. A GPU processes the recording, and the app returns the transcript and measurements for a teacher to review.

  1. 1

    Record a grade- and language-matched passage

  2. 2

    Upload audio directly to private R2 storage

  3. 3

    Dispatch a job from the Worker

  4. 4

    Detect speech

  5. 5

    Transcribe with the empirically selected Whisper configuration

  6. 6

    Align words and time

  7. 7

    Compare the transcript with the reference passage

  8. 8

    Compute recording-level raw measurements

  9. 9

    Calibrate with required grade and language features

  10. 10

    Store results through the Worker API

  11. 11

    Present evidence for teacher review

Passages used in the study

The partner school selected the two English and two Filipino Phil-IRI passages used at each grade. These Set 1 and Set 2 study identifiers are not the public Demo/Practice A–D catalogue.

Grade 4

English: Set 1 — The Best Part of the Day; Set 2 — Ice Cream for Sale

Filipino: Set 1 — Ang Aso sa Lungga; Set 2 — Si Muning!

Grade 5

English: Set 1 — Early Start; Set 2 — Rice for Lunch

Filipino: Set 1 — Ang Regalo kay Lea; Set 2 — Balat ng Saging

Grade 6

English: Set 1 — Chameleons; Set 2 — The Philippine Eagle

Filipino: Set 1 — Paano Nabubuo ang Isang Batas?; Set 2 — Panahon ng Bagong Bato

Phil-IRI Manual 2018 — DepEd Learning Resource Management and Development System

Validation

We compare model sizes by word error rate (WER), compare raw and calibrated scores, and test language configurations. For continuous measurements, we assess agreement using the intraclass correlation coefficient (ICC) and Bland–Altman analysis.