Turn an Agent into a robot that can understand voice and reply in a persona voice—the first IM voice duplex closed-loop skill.
- 🎙 Understands voice: any voice file to text,supports WeChat .silk format(unique—most open-source solutions cannot decode WeChat voice)
- 🔊 Voice bubble replies: persona reply → Edge TTS multi-voice → silk-encoded voice sent back to WeChat(pilk bidirectional, first in the industry to achieve this)
- 🗣 Persona replies: 6 built-in personas (Energetic Girl / Gentle Confidante / Steady Uncle / News Anchor / Sunny Buddy / Casual Friend), each with a dedicated Chinese voice + tone
- 💰 Zero API cost: faster-whisper (local) + Edge TTS (free) + pilk (open source) all free, optional LLM to enhance tone
Voice chat for agents: transcribe voice (incl. WeChat silk) → reply with persona voices → back to WeChat voice bubbles. First IM two-way voice skill.
Differences from similar solutions (tested research as of 2026-09)
| Solution | Problem |
|---|---|
| jozhn/wechat-voice-decode-skill | Only one-way decoding and transcription, no TTS/persona/voice reply, locked to the openclaw path |
| zhayujie/chatgpt-on-wechat (46k★) | Voice relies on cloud ASR/TTS, personal WeChat channel has account ban risk, no native silk closed loop |
| whisperbot / Telegram transcription bot family | All one-way "voice→text", no multi-voice persona voice replies |
| SillyTavern (33k★) | Character voices are powerful but only in their own UI, not connected to WeChat/Telegram voice messages |
| Microsoft cloud voice skills (Azure/OpenAI TTS) | Single capability + paid cloud API, no IM scenario |
Others do "dictation" or "read-aloud"; voice-persona does "speak with a persona voice in WeChat/Telegram".
Quick validation / Smoke Test (30 seconds, no key required)
Run from this skill directory (this repository already includes the scripts, no extra clone needed):
cd plugins/all-skills/skills/voice-persona # 或直接进入 voice-persona 技能目录
python3 scripts/voice_persona.py demo
# 🎙 语音 → 文字 → [元气少女/沉稳大叔/新闻播报] 三音色回复音频
# ✅ 全链路通过：语音输入 → 转写 → 人格回复 → 语音输出
Prerequisites (dependencies)
pip install faster-whisper pilk edge-tts # 首次使用安装
# 可选：--llm 增强口吻需要 LLM_API_KEY / OpenAI 兼容 URL（与本集合其他技能一致）
Usage
All commands call the scripts from within the skill directory using relative paths (no need to install to system PATH):
# 1. 语音 → 文字（微信语音直接传 .silk 文件即可）
python3 scripts/voice_persona.py stt wechat_voice.silk
python3 scripts/voice_persona.py stt meeting.m4a --model small
# 2. 文字 → 人格回复 → 语音 mp3
python3 scripts/voice_persona.py speak "明天记得交报告" --persona yunjian --out reply.mp3
# 3. 只取人格化文本（接你自己的 TTS/IM）
python3 scripts/voice_persona.py chat "明天记得交报告" --persona xiaoyi
# 4. 列出人格
python3 scripts/voice_persona.py list
# 5. 音频 → 微信 silk 语音（可回发微信语音气泡）
python3 scripts/voice_persona.py to_silk reply.mp3 --out reply.silk
# 6. 一条命令双向闭环：人格语音 → 微信格式
python3 scripts/voice_persona.py speak "明天早上十点开会" --persona xiaoyi --out r.mp3
python3 scripts/voice_persona.py to_silk r.mp3 # → r.silk（#!SILK_V3）
Install in Hermes (optional)
This repository's scripts can be run directly; if using Hermes Agent, you can also install from the skill source repository for automatic integration:
hermes skills install jiawood2006/hermes-skills/skills/voice-persona
Persona library (extensible)
In voice_persona.py at the top PERSONAS / STYLE_WORDS just add:
| key | Persona | Edge TTS voice | Tone |
|---|---|---|---|
| xiaoyi | Energetic Girl · Xiaoyi | zh-CN-XiaoyiNeural | hehe, la/ya |
| xiaoxuan | Gentle Confidante · Xiaoxuan | zh-CN-XiaoxuanNeural | don't worry, take your time |
| yunjian | Steady Uncle · Yunjian | zh-CN-YunjianNeural | speak plainly, conclusion first |
| yunyang | News Anchor · Yunyang | zh-CN-YunyangNeural | broadcast, well-organized |
| yunxi | Sunny Buddy · Yunxi | zh-CN-YunxiNeural | come on, no problem |
| xiaochen | Casual Friend · Xiaochen | zh-CN-XiaochenNeural | colloquial, like a friend |
Agent integration (Hermes and other frameworks)
STT integration: pass the voice message file received by the platform to stt subcommand → get the text into the conversation stream.
Hermes's stt.provider: local_command + HERMES_LOCAL_STT_COMMAND environment variable can directly connect WeChat voice automatically (pointing to the absolute path of the script under this skill's scripts/):
HERMES_LOCAL_STT_COMMAND=<python> <...>/skills/voice-persona/scripts/voice_persona.py stt {input_path} --model {model} --output_dir {output_dir} --language {language}
TTS integration: personalized text → speak outputs mp3 → the platform sends voice.
Known pitfalls
- WeChat silk must be decoded with pilk：
pilk.decode()outputs raw PCM without a RIFF header, which whisper cannot read—usepilk.silk_to_wav()(outputs standard wav, already tested) - First run faster-whisper will download the model (small ~500MB), waiting 1-3 minutes is normal
- Edge TTS requires internet(Microsoft's free interface); when offline, speak will fail, stt/chat are unaffected
- Transcription quality: on Intel Mac CPU with int8, the small model is good enough for Chinese; for noisy/dialect audio, base/small is recommended, or resend
Verification record (real output)
- WeChat .silk voice transcription tested successfully ("API这一块我会让海沃云来对接…" 39 characters/17 characters, two segments)
- demo full pipeline: say synthesizes Chinese → stt transcribes → 3 persona replies → 3 mp3s (27-31KB) ✅