← All work

MasterMyAccent

Being understood at work and on the phone, not sounding native.

Being rebuilt as a small pronunciation course for Amharic speakers working in the US and Canada. First, a teacher's blind marking tests whether Azure's scoring hears their mistakes.

The idea

MasterMyAccent is for Amharic speakers working in the US and Canada who need to be understood in interviews, at work and on the phone. Many pronunciation apps chase a native accent and hand out confident scores without showing they can hear the mistakes this group actually makes. Here an accent is not an error. The goal is clear speech on the specific sounds that cause trouble.

The planned course is six short units built around a few trouble spots, such as th or the vowels in ship and sheep. A learner would read a sentence aloud, see which sound got in the way, hear a model, and try again. The bet is to buy the scoring from Azure instead of building an engine, and to keep a teacher in the loop, so the effort goes into serving one narrow group well.

Nothing is promised before it is checked. Automatic feedback will cover only the trouble spots where Azure's flags agree with a teacher's blind marking, against a bar fixed before any results; every other spot stays teacher-led. Today the 2025 app is retired with sign-ups paused, the engine-check kit is built with its rules locked, and no participant has been recorded yet.

How it works

  1. Record
  2. Azure assessment
  3. Teacher comparison
  4. Automatic or teacher-led feedback

A Next.js app with Supabase auth and Postgres. The retired path sends recordings to a separate phoneme-timing service. The pilot path sends 16 kHz clips to Azure Pronunciation Assessment, and a laptop-only recorder and scripts run the engine check.

Where it stands today

The 2025 app was retired on 23 September 2026 because its engine timed sounds but could not tell whether the right sound was said. It is being revived for Amharic speakers working in the US and Canada, with scoring bought from Azure Pronunciation Assessment. The engine-check kit is merged, and its scoring rules were hash-locked before any results; no participant has been recorded yet. The live site still runs the retired timing-only app and says new sign-ups are paused.

Working now

  • A laptop-only recorder that saves 16 kHz clips with the microphone processing the browser actually applied
  • Azure Pronunciation Assessment returning phoneme scores, spoken-phoneme alternatives, and prosody, tested on synthetic voices only
  • Engine-check scoring rules hash-locked before any results, a blind teacher-sheet generator, and a withdrawal script
  • A planted-mistake check on synthetic voices showing that each trouble spot's measure can fire
  • The retired app, still live with sign-ups paused: timing-only analysis, written summaries, and deletion of recordings and accounts

Not there yet

  • Whether Azure hears these learners' mistakes is still unanswered. Sign-in email delivery, payments, and a legal review of voice recordings must be done before any learner is invited.

How it got here

  1. A separate voice-agent accent-tutor prototype came first
  2. MasterMyAccent's own codebase started
  3. A separate phoneme-timing service built for pronunciation analysis
  4. Named MasterMyAccent and broadened to non-native speakers of American English
  5. Development paused
  6. Hardening began after a September audit found 185 confirmed issues
  7. Retired, because its engine timed sounds but could not tell whether the right sound was said; revived the same day with a new thesis
  8. Engine-check kit merged and the honest site update live; scoring rules locked before any results
  9. Pre-registration revision 2 locked on its branch; it merges once the repository's automated checks can run again

What comes next

Next step

Record the owner's clips to fix the microphone setting and transport, then run the pre-registered engine check with 15–20 participants and a teacher's blind marking.

Later ideas, not current features

  1. Record the owner's clips to choose the microphone setting and transport
  2. Record 15–20 consenting adult Amharic speakers, and have a teacher mark the clips blind
  3. Score once against the locked bar
  4. Build a six-unit course with email-code sign-in; spots that fail stay teacher-led
  5. Run an invite-only pilot for 10–20 learners

What I checked, and what stays bounded

Dated evidence

  1. Retirement record and revival decisiondecision
  2. Live site says sign-ups are paused and describes timing-only analysisdeployment
  3. Engine-check rules hash-locked before any results (revision 2, on its branch)pre-registration

Boundaries it keeps

  • No course exists yet, and no learner has used the pilot
  • No participant has been recorded; Azure has been tested only on synthetic voices
  • No automatic feedback is promised for a trouble spot unless it passes the pre-registered bar on held-out speakers; spots that fail stay teacher-led
  • No payments are taken; billing is off
  • Live voice coaching is disabled and will be removed
  • No objective scoring or learner-improvement claim
  • Participants' audio and marks are never public; only totals are reported

My role

Chose the audience, the decision to buy scoring rather than build it, and the engine-check bar; reviewed each plan revision and rule, and approved every merge and production change.