How tone scoring actually works

The 4-step flow

1. 1

   You record

   Your mic captures ~2 seconds of audio in your browser. No app install. The audio goes to Maa over HTTPS — we store it in Supabase Storage (encrypted at rest, Singapore ap-southeast-1 region for low Thai-user latency).

2. 2

   Whisper transcribes

   OpenAI Whisper turns your audio into Thai script + syllable timestamps. We use Whisper-1 — same model that powers ChatGPT voice, plenty good at Thai.

3. 3

   Claude grades

   Anthropic Claude Sonnet 4.6 compares each syllable you produced against the expected tone (mid / low / falling / high / rising). It rates the whole phrase 0–100 and flags every off-tone syllable individually.

4. 4

   You see feedback

   A score card pops up in 2-4 seconds with per-syllable check marks, an English line of feedback, and a Thai line so you start absorbing the vocabulary of tone-correction itself. Hit retry to try again — free tier = 50 attempts/day.

What each score range means

• 90–100ปังมาก 🔥

  Locals will get it instantly. Move on.

• 70–89เก่งจัง ✨

  Solid — minor wobble on 1 syllable. Replay for fun, don't sweat it.

• 50–69ใช้ได้

  Understandable, but at least one tone is off enough to maybe confuse fast speech. Listen + try again.

• 0–49ลองอีกที 🎧

  Tone(s) flipped. Listen to the model audio, then re-record. This is the most useful score — it tells you exactly which syllable to focus on.

What scoring measures

• Tone direction per syllable — the main event. Thai has 5 tones; getting them right is ~80% of being understood.
• Syllable count — did you say all the syllables, in order?
• Recognizable consonants + vowels — Whisper has to be able to transcribe what you said.

What it does not measure

• Accent or “native-ness” — Maa is calibrated for foreign learners. We're strict on tone errors and generous on accent.
• Speed or rhythm — not yet. Roadmap item.
• Politeness particles (ครับ / ค่ะ) — separate lesson topic, not part of tone scoring.
• Whether your phrase is socially appropriate — that's in the lesson's cultural notes, not the score.

Your audio + privacy

Recordings are stored for up to 30 days for AI improvement, then auto-deleted. You can delete all of yours instantly from Settings (button: “Delete my recordings now”). We don't use your voice to train models for other users.

Full details in Privacy.

When the AI gets it wrong

It happens. Background noise, mic compression, or a regional accent can confuse Whisper. The retry button is there for a reason — the system does not penalize you for re-attempts within your daily quota.

If you feel a score is consistently unfair, email hi@maa.app with the lesson + your audio (we'll find it in your account) and we'll tune the rubric.