Article Tts

中文 OCR 语言包安装：

• Linux（WSL/Debian/Ubuntu）：apt-get install tesseract-ocr-chi-sim
• macOS：brew install tesseract-lang（自带中文）
• Windows：下载 chi_sim.traineddata 放入 Tesseract 安装目录的 tessdata 文件夹

Workflow

Input Types

• 图片：OCR 提取文字（需要 lang 指定语言）
• 纯文字：直接 TTS，无需 OCR

Standard Flow（默认，需确认）

图片 → OCR 提取文字 → 展示给用户确认 → 用户确认 → 生成 TTS → 发送
文字 → 直接生成 TTS → 发送

Skip-Confirmation Flow ⚠️

用户说"不需要确认"或"直接生成"时，跳过确认步骤。

⚠️ 安全提示：skipConfirmation 会跳过文字确认步骤，OCR 提取的文本（可能包含敏感信息）会直接转为音频并发送。适用于可信来源、低敏感内容。建议默认关闭（skipConfirmation: false）。

OCR Step

# 图片预处理
from PIL import Image, ImageOps
img = Image.open(image_path)
img = ImageOps.autocontrast(img.convert('L'), cutoff=10)
w, h = img.size
img = img.resize((w*4, h*4), Image.LANCZOS)
img.save('/tmp/ocr_input.jpg', quality=99)

# 英文
tesseract /tmp/ocr_input.jpg stdout -l eng --psm 4

# 中文
tesseract /tmp/ocr_input.jpg stdout -l chi_sim --psm 4

TTS Step

全文字频

uvx edge-tts \
  -t "FULL TEXT" \
  -v en-US-EmmaNeural \
  --rate=-10% \
  --write-media OUTPUT_DIR/full_article.mp3

# 中文
uvx edge-tts \
  -t "中文文字内容" \
  -v zh-CN-XiaoxiaoNeural \
  --rate=-10% \
  --write-media OUTPUT_DIR/full_article.mp3

按句拆分（仅 splitSentences=true）

import subprocess, re

def split_sentences(text, lang='en'):
    if lang == 'zh':
        # 中文按句号/感叹号/问号拆分
        sentences = re.split(r'(?<=[。！？])\s*', text)
    else:
        # 英文按 .!? 拆分
        sentences = re.split(r'(?<=[.!?])\s+', text)
    return [s.strip() for s in sentences if s.strip()]

sentences = split_sentences(text, lang=lang)
for i, sentence in enumerate(sentences, 1):
    num = str(i).zfill(2)
    voice = 'zh-CN-XiaoxiaoNeural' if lang == 'zh' else 'en-US-EmmaNeural'
    subprocess.run([
        "uvx", "edge-tts",
        "-t", sentence,
        "-v", voice,
        "--rate=-10%",
        "--write-media", f"OUTPUT_DIR/sentence_{num}.mp3"
    ])

Output Directory

/mnt/d/wslspace/workspace/articles/YYYY-MM-DD-article-slug/
├── original_text.md
├── full_article.mp3
└── sentence_01.mp3 ...

Sending via Message Channel

The agent detects the active channel from the runtime context and calls message(...) accordingly. No hardcoded channel — the agent uses whichever channel the user is currently chatting through.

# Detect active channel automatically (from runtime inbound metadata)
# channel is inferred: feishu / telegram / discord / whatsapp / signal / imessage / openclaw-weixin

# 发送全文
message(action="send", channel="{active_channel}",
        message="📄 全文音频",
        media="PATH/full_article.mp3",
        filename="full_article.mp3")

# 发送每句
for i, sentence in enumerate(sentences, 1):
    num = str(i).zfill(2)
    message(action="send", channel="{active_channel}",
            message=f"📝 {num}: {sentence}",
            media=f"PATH/sentence_{num}.mp3",
            filename=f"sentence_{num}.mp3")

Channel Behavior Notes

If the channel does not support audio, the agent saves the file to OUTPUT_DIR and sends the file path as a text message instead.

Available TTS Voices

English

en-US-EmmaNeural, en-US-BrianNeural, en-GB-LibbyNeural, ...

Chinese

zh-CN-XiaoxiaoNeural（女声）, zh-CN-YunxiNeural（男声）, zh-CN-YunyangNeural（新闻男声）, ...

查看完整列表：uvx edge-tts -l | grep "zh-CN"

Notes

• Tesseract + English 预装；中文需 apt-get install tesseract-ocr-chi-sim
• edge-tts 通过 uvx 运行，无需安装
• 图片质量直接影响 OCR 效果，尽量保持光线充足、角度端正