edge-tts in Practice: Give Your Content Site a Voice

Using free edge-tts to add audio functionality to a content site — SFD Lab practical tutorial including multilingual support, batch generation pipeline, pitfall notes, and comparison with openai-tts.

The Use Case

SFD Lab runs smallfiredragon.com, a content site with articles in three languages (zh-CN, en, zh-TW). We wanted to add audio versions of articles — but at scale, paying per-character TTS adds up quickly.

edge-tts runs free on our servers and covers all three languages with natural-sounding voices.

The Stack

edge-tts (Microsoft Neural TTS via Edge browser API)
OpenClaw with the edge-tts skill installed
A simple cron job to auto-generate audio when articles are published

Installation

clawhub install edge-tts
pip install edge-tts  # Python library also required for batch generation

Batch Generation Pipeline

import asyncio
import edge_tts

async def generate_article_audio(text, locale, output_path):
    voice_map = {
        'zh-CN': 'zh-CN-XiaoxiaoNeural',
        'en': 'en-US-JennyNeural',
        'zh-TW': 'zh-TW-HsiaoChenNeural'
    }
    voice = voice_map.get(locale, 'zh-CN-XiaoxiaoNeural')
    communicate = edge_tts.Communicate(text, voice)
    await communicate.save(output_path)

# Generate for all three locales
async def main():
    tasks = [
        generate_article_audio(zh_text, 'zh-CN', 'article_zh.mp3'),
        generate_article_audio(en_text, 'en', 'article_en.mp3'),
        generate_article_audio(tw_text, 'zh-TW', 'article_tw.mp3'),
    ]
    await asyncio.gather(*tasks)

asyncio.run(main())

Voice Quality Comparison

Provider	Quality	Cost	Languages	Latency
edge-tts	Good	Free	74	Medium
openai-tts	Excellent	$15/M chars	Limited	Fast
Google TTS	Excellent	$4/M chars	220+	Fast

For a content site where audio is a supplementary feature (not the primary product), edge-tts quality is entirely sufficient. We'd switch to openai-tts only if edge-tts became unavailable.

Production Pitfalls

Microsoft rate limiting: In batch mode, Microsoft will rate-limit if you fire too many concurrent requests. Add delays between requests: 2-3 seconds between each generation.

Long article chunking: Articles over 3000 characters need to be split into paragraphs, generated separately, and stitched together with ffmpeg. We built a simple chunking function that splits on paragraph breaks.

Voice consistency: If you regenerate part of an article, make sure to use the same voice ID. Mixing voices mid-article is jarring.

Caching strategy: Generate audio once and cache it. Re-generating on every page load would hit rate limits fast.

The OpenClaw Integration

With the edge-tts skill installed in OpenClaw, you can trigger audio generation conversationally:

Generate audio for this article in English using en-US-JennyNeural
[paste article text]

The skill handles chunking, generation, and returns the audio file path. Useful for one-off generation or when you want to preview a specific voice before committing to batch generation.

Summary

edge-tts is the right choice for any content site that wants audio without paying per-character. Zero cost, reasonable quality, 74 languages. The only investment is the setup time for a proper batch generation pipeline.

SFD Lab | edge-tts production deployment notes