name	english-pronunciation-audio
description	Generate TTS audio for English practice replies and send via Telegram. Use together with english-daily-coach. Extracts spoken English lines, skips Chinese, and sends one audio file per reply. When user asks "怎么读", generate audio for the English sentence — do NOT explain pronunciation in text.

English Pronunciation Audio

CRITICAL Rules

This skill generates AUDIO files only. NEVER output text-based pronunciation guides (e.g. "vuh-LOR-unt", IPA symbols).
When user asks "X怎么读" (how to pronounce the sentence they sent): Pass ONLY the exact sentence X the user gave you. Example: user says "What would you like to order for lunch today?怎么读" → exec with --text "What would you like to order for lunch today". Do NOT pass your reply or any extra text.
When sending coaching reply: Pass the full draft (你说/➡️/💬/📚/🎯) so the script extracts translation + examples + question.
Generate at most one audio file per reply.
Skip Chinese lines (lines starting with 提示:, 你说:, 📚, 💬).
Skip You said: lines. Only read More natural: and Reusable phrase: lines.
Skip vocabulary definitions (• word — 释义) and phonetic lines (IPA, "pronounced", "vuh-LOR-unt").
If the user ends the session, skip audio for the final reply.

Supported Reply Shapes

Spoken content is extracted from:

➡️ line (translation)
①②③ lines (example sentences)
Line after 🎯 我会这样继续问你： (follow-up question)
Translation: / You can say: / Now you try: / More natural: / Reusable phrase: (legacy labels)

Script

Use scripts/tts_openrouter.py.
Pass the full draft reply text with --text. The script extracts spoken English automatically.
Config: assets/tts-config.json.
If script returns no_spoken_text or too_long, send text reply only.
If script fails, do not block the text reply. Retry at most once.

Reply Shaping

Keep extracted English under 300 characters total.
Keep Chinese on separate 提示: lines.
One template per reply for deterministic extraction.

Debugging

--extract-only: inspect extracted English as JSON.
--dry-run: generate audio but skip Telegram upload.
To verify what audio will be sent: python3 scripts/tts_openrouter.py --text "your text" --extract-only

English Pronunciation Audio

CRITICAL Rules

This skill generates AUDIO files only. NEVER output text-based pronunciation guides (e.g. "vuh-LOR-unt", IPA symbols).

When user asks "X怎么读" (how to pronounce the sentence they sent): Pass ONLY the exact sentence X the user gave you. Example: user says "What would you like to order for lunch today?怎么读" → exec with --text "What would you like to order for lunch today". Do NOT pass your reply or any extra text.

When sending coaching reply: Pass the full draft (你说/➡️/💬/📚/🎯) so the script extracts translation + examples + question.

Generate at most one audio file per reply.

Skip Chinese lines (lines starting with 提示:, 你说:, 📚, 💬).

Skip You said: lines. Only read More natural: and Reusable phrase: lines.

Skip vocabulary definitions (• word — 释义) and phonetic lines (IPA, "pronounced", "vuh-LOR-unt").

If the user ends the session, skip audio for the final reply.

Supported Reply Shapes

Spoken content is extracted from:

➡️ line (translation)

①②③ lines (example sentences)

Line after 🎯 我会这样继续问你： (follow-up question)

Translation: / You can say: / Now you try: / More natural: / Reusable phrase: (legacy labels)

Script

Use scripts/tts_openrouter.py.

Pass the full draft reply text with --text. The script extracts spoken English automatically.

Config: assets/tts-config.json.

If script returns no_spoken_text or too_long, send text reply only.

If script fails, do not block the text reply. Retry at most once.

Reply Shaping

Keep extracted English under 300 characters total.

Keep Chinese on separate 提示: lines.

One template per reply for deterministic extraction.

Debugging

--extract-only: inspect extracted English as JSON.

--dry-run: generate audio but skip Telegram upload.

To verify what audio will be sent: python3 scripts/tts_openrouter.py --text "your text" --extract-only

english-pronunciation-audio