JLPT N5 Listening Select Audio Tester (Mondai 3 / 発話表現)

If any FAIL: Stop. Report errors. Fix the JSON. Re-run Pass 1. If PASS: Proceed to Pass 2.

Pass 2 — Semantic / Linguistic Review (Claude)

Read derived-data.json and check each criterion:

2.1 — N5 Level Appropriateness

• All text in transcriptions.ja.intro, options[].text, and question must use N5 vocabulary and grammar.
• No N4+ vocabulary. No complex conjugations beyond N5.
• Check: Any word or pattern above N5?

2.2 — Correct Option Validity

• transcriptions.ja.options[correct_option].text must be the natural, appropriate thing for the arrow character to say in the described situation.
• Check: Is the marked correct option actually the best answer for the scenario?

2.3 — Distractor Plausibility (Trap Quality)

• Each wrong option must use one of the 4 trap types: Rol Karışıklığı, Yön Karışıklığı, Register Karışıklığı, Bağlam Karışıklığı.
• Distractors must be plausible enough to trick a careless learner.
• Check: Does each distractor have a clear, identifiable trap type? Is it plausible?

2.4 — Options Naturalness

• All 3 Japanese options must be natural utterances.
• Options should be parallel in structure and speech register.
• Intro must correctly describe the situation in N5 Japanese.
• Check: Does every option sound like something a real person would say?

2.5 — Translation Accuracy (6 Languages)

• transcriptions.tr/en/de/fr/es/ko must accurately reflect the Japanese intro, options, and question.
• Speaker role (customer, student, etc.) must be preserved in translation.
• Check: Are all 6 translations accurate and natural?

2.6 — Logic Explanation Quality

• logic.tr and logic.en must explain why the correct option is right.
• Must explain why each distractor is wrong (name the trap type).
• Check: Are the logic explanations accurate, complete, and clear?

Output Format for Pass 2

Pass 2 — Semantic Review for {id}
  2.1 N5 Level:           ✅ PASS / ❌ FAIL (details)
  2.2 Correct Option:     ✅ PASS / ❌ FAIL (details)
  2.3 Distractor Quality: ✅ PASS / ❌ FAIL (trap types identified)
  2.4 Naturalness:        ✅ PASS / ⚠️ WARN (details)
  2.5 Translations:       ✅ PASS / ❌ FAIL (details)
  2.6 Logic Quality:      ✅ PASS / ❌ FAIL (details)

If any 2.1–2.3 FAIL: Stop. Report issues. Suggest fixes. Re-run after fixing. If PASS or only WARNs: Proceed to Pass 3.

Pass 3 — TTS Script Generation

Generate tts_script.json in the same folder.

Speaker Voice Rules

Determine the arrow character's gender by reading the image (or ask the user if ambiguous):

TTS Script Structure

{
  "tts_script": [
    { "voice": "Intro_Voice", "text": "<transcriptions.ja.intro>" },
    { "break": "1s" },
    { "voice": "<Speaker_Voice>", "text": "１　<options[0].text>" },
    { "break": "0.8s" },
    { "voice": "<Speaker_Voice>", "text": "２　<options[1].text>" },
    { "break": "0.8s" },
    { "voice": "<Speaker_Voice>", "text": "３　<options[2].text>" },
    { "break": "1s" },
    { "voice": "Intro_Voice", "text": "<transcriptions.ja.question>" }
  ]
}

Rules

1. Intro_Voice reads the intro and the question repeat.
2. 1s break after intro.
3. The same Speaker_Voice (Male_1 or Female_1) reads ALL 3 options — they all come from the same character.
4. Prepend each option text with １　, ２　, ３　 (full-width digits + space).
5. 0.8s break between each option.
6. 1s break before question repeat.
7. Text must be Japanese only.
8. Every entry has EITHER {voice, text} OR {break}, never both.

Save to: backend/listening/data/selectAudio/{id}/tts_script.json

After saving: Display the script to the user.

Pass 4 — Image Generation (Imagen 3)

NOTE: The arrow is NOT added by this pass. The user adds it manually after image generation.

Step 4.1 — Read image_prompt

Read image_prompt from derived-data.json.

Step 4.2 — Generate image.png

STYLE RULE (MANDATORY): All images must use the approved manga/textbook style. The image_prompt in derived-data.json should already follow this style. If it doesn't, rewrite it to include these required keywords:

• Simple manga-style line drawing, black and white
• simple round head (for each character)
• Cartoon style with simple round faces
• clean outlines, white background, no shading, no color, no text, no arrows
• Same style as JLPT N5 textbook manga illustrations

AVOID: JLPT exam textbook line drawing → produces realistic/Western faces (wrong style).

Use the generate_image script pattern adapted for selectAudio:

python3 -c "
import os, json, base64
from pathlib import Path