davepoon/buildwithclaude / plugins/all-skills/skills/voice-persona
voice-persona
Verified让 Agent 变成能语音对话的机器人:语音文件转文字(支持微信 silk 格式)+ 多音色人格回复(Edge TTS 免费中文音色),全本地零 API 成本。Voice persona chat: transcribe voice messages (incl. WeChat silk) and reply in persona voices.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 5.7 KB |
| scripts/voice_persona.py | script | 10.9 KB |
scripts/voice_persona.py
View on GitHub#!/usr/bin/env python3
"""
voice-persona — 给 Agent 装上"语音对话 + 人格音色"(Voice persona chat for agents)
================================================================
能力:
1. stt — 任意语音文件转文字(含微信 silk,pilk 解码;其余格式直接喂 faster-whisper)
2. speak — 把文字变成"某个人格"的口吻,并用 Edge TTS 免费中文音色合成语音
3. chat — 端到端:给文字 → 人格化回复文本
4. demo — 免 key 全链路验证:合成一句输入 → 人格回复 → 输出音频
人格 = 音色 + 说话风格。内置 6 个人格(见 PERSONAS),可自行扩展。
依赖(都是免费/本地):
pip install faster-whisper pilk edge-tts # stt 需要前两个,speak 需要 edge-tts
用法示例:
python3 voice_persona.py stt wechat_voice.silk # 微信语音 → 文字
python3 voice_persona.py speak "明天记得交报告" --persona xiaoyi --out reply.mp3
python3 voice_persona.py chat "明天记得交报告" --persona dajiang
python3 voice_persona.py demo # 全链路自测
"""
import argparse, os, re, sys, tempfile, pathlib, subprocess, shutil
# ═══════════════ 人格库(可自行增删) ═══════════════
# persona: 音色id | 语气前缀 | 风格词(用于LLM可选增强) | 示例开场
PERSONAS = {
"xiaoyi": {"voice": "zh-CN-XiaoyiNeural", "style": "元气少女", "emoji": "✨",
"prefix": "嘿嘿,",
"desc": "元气少女·小伊:活泼可爱,爱用'啦/呀/喔'"},
"xiaoxuan": {"voice": "zh-CN-XiaoxuanNeural", "style": "温柔知心", "emoji": "💗",
"prefix": "嗯嗯,别担心,",
"desc": "温柔知心·晓萱:柔和、安抚、耐心"},
"yunjian": {"voice": "zh-CN-YunjianNeural", "style": "沉稳大叔", "emoji": "📌",
"prefix": "好,直说:",
"desc": "沉稳大叔·云健:简洁、稳重、结论先行"},
"yunyang": {"voice": "zh-CN-YunyangNeural", "style": "新闻播报", "emoji": "📰",
"prefix": "播报一则消息:",
"desc": "新闻播报·云扬:正式、条理、有节奏"},
"yunxi": {"voice": "zh-CN-YunxiNeural", "style": "阳光伙伴", "emoji": "☀️",
"prefix": "没问题!",
"desc": "阳光伙伴·云希:热情、鼓励、正能量"},
"xiaochen": {"voice": "zh-CN-XiaochenNeural", "style": "随性好友", "emoji": "🎮",
"prefix": "行,跟你说——",
"desc": "随性好友·晓辰:口语化、像朋友聊天"},
}
STYLE_WORDS = {
"xiaoyi": ["啦", "呀", "超", "好好玩"],
"xiaoxuan": ["别担心", "慢慢来", "可以的", "我理解"],
"yunjian": ["首先", "其次", "重点是", "结论是"],
"yunyang": ["各位听众", "值得注意的是", "总而言之"],
"yunxi": ["加油", "真棒", "一起搞定", "没问题"],
"xiaochen": ["讲真", "就是说", "咱", "差不多"],
}
# ═══════════════ 一、STT:语音 → 文字 ═══════════════
def _decode_silk(src: str) -> str:
"""silk → 标准 wav 临时文件;非 silk 原样返回"""
if pathlib.Path(src).suffix.lower() != ".silk":
return src
import pilk
tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False); tmp.close()
pilk.silk_to_wav(src, tmp.name) # 24kHz/16bit/mono 标准 wav
return tmp.name
def stt_file(src: str, model: str = "small", language: str = "zh") -> str:
from faster_whisper import WhisperModel
wav = _decode_silk(src)
try:
m = WhisperModel(model, device="cpu", compute_type="int8")
segs, _info = m.transcribe(wav, language=language or None)
return "".join(s.text for s in segs).strip()
finally:
if wav != src:
try: os.unlink(wav)
except OSError: pass
# ═══════════════ 二、人格化:文字 → 人格口吻 ═══════════════
def personalize(text: str, persona: str, use_llm: bool = False) -> str:
key = persona if persona in PERSONAS else "xiaoyi"
p = PERSONAS[key]
if use_llm:
try:
return _llm_personalize(text, key, p)
except Exception:
pass # LLM 失败自动降级规则
return p["prefix"] + text
def _llm_personalize(text: str, key: str, p: dict) -> str:
import json, urllib.request
llm_key = os.environ.get("LLM_API_KEY", "").strip()
if not llm_key:
raise RuntimeError("no LLM key")
style_words = STYLE_WORDS.get(key, [p["style"]])
prompt = (f"把下面这段话改写成{p['desc']}的口吻(自然口语、简短,保留原意,"
f"可适当用风格词如{'/'.join(style_words[:3])}),"
f"只输出改写后的内容,不要解释:\n{text}")
body = json.dumps({"model": os.environ.get("LLM_MODEL", "deepseek-chat"),
"messages": [{"role": "user", "content": prompt}],
"temperature": 0.8, "max_tokens": 300}).encode()
req = urllib.request.Request(os.environ.get("LLM_BASE_URL", "https://api.deepseek.com/v1/chat/completions"),
data=body, headers={"Content-Type": "application/json", "Authorization": f"Bearer {llm_key}"})
with urllib.request.urlopen(req, timeout=60) as r:
d = json.load(r)
return d["choices"][0]["message"]["content"].strip()
# ═══════════════ 三、TTS:文字 → 语音 ═══════════════
def speak(text: str, persona: str, out: str = "") -> str:
p = PERSONAS.get(persona) or PERSONAS["xiaoyi"]
out = out or os.path.join(tempfile.gettempdir(), f"voice_persona_{persona}.mp3")
import edge_tts
async def _run():
com = edge_tts.Communicate(text, p["voice"])
await com.save(out)
import asyncio
asyncio.run(_run())
return out
# ═══════════════ 四、silk 编码:语音回发微信(双向闭环) ═══════════════
def to_silk(audio: str, out: str = "") -> str:
"""任意音频(mp3/wav/m4a) → 微信 silk 语音(#!SILK_V3 头),可回发微信语音气泡。
解码优先 ffmpeg,无则用 macOS 原生 afconvert;均转 24kHz 单声道 pcm。"""
import pilk, subprocess, tempfile, shutil, wave
out = out or str(pathlib.Path(audio).with_suffix(".silk"))
pcm = tempfile.NamedTemporaryFile(suffix=".pcm", delete=False); pcm.close()
tmpwav = None
try:
ffmpeg = shutil.which("ffmpeg") or "/usr/local/bin/ffmpeg"
if os.path.exists(ffmpeg):
r = subprocess.run([ffmpeg, "-y", "-v", "error", "-i", audio,
"-f", "s16le", "-ac", "1", "-ar", "24000", pcm.name],
capture_output=True)
if r.returncode != 0:
return f"ffmpeg error: {r.stderr.decode()[:200]}"
else:
tmpwav = tempfile.NamedTemporaryFile(suffix=".wav", delete=False); tmpwav.close()
r = subprocess.run(["/usr/bin/afconvert", "-f", "WAVE", "-d", "LEI16@24000",
"-c", "1", audio, tmpwav.name], capture_output=True)
if r.returncode != 0:
return f"afconvert error: {r.stderr.decode()[:200]}"
with wave.open(tmpwav.name, "rb") as w:
raw = w.readframes(w.getnframes())
with open(pcm.name, "wb") as f:
f.write(raw)
pilk.encode(pcm.name, out, pcm_rate=24000, tencent=True)
return out
finally:
try: os.unlink(pcm.name)
except OSError: pass
if tmpwav:
try: os.unlink(tmpwav.name)
except OSError: pass
# ═══════════════ demo:免 key 全链路自测 ═══════════════
def cmd_demo():
print("🧪 voice-persona 自测(全本地免费,无需 key)")
# 1) 合成一段中文输入当"语音消息"
say_path = os.path.join(tempfile.gettempdir(), "vp_input.m4a")
try:
subprocess.run(["say", "-v", "Ting-Ting", "-o", say_path, "明天早上十点开会记得提前准备材料"],
check=True, capture_output=True, timeout=60)
except Exception:
print("⚠️ 跳过 say 合成输入(当前系统无中文语音)")
return
# 2) STT 转文字
txt = stt_file(say_path)
print(f"🎙 语音 → 文字: {txt}")
# 3) 三种人格演示
for pname in ["xiaoyi", "yunjian", "yunyang"]:
p = PERSONAS[pname]
reply = personalize(txt, pname)
audio = speak(reply, pname)
print(f"🗣 [{p['style']}] {reply}")
print(f" 🔊 {audio} ({os.path.getsize(audio)//1024}KB)")
# 4) 双向闭环:人格语音 → 微信 silk
audio_yunyang = speak(personalize(txt, "yunyang"), "yunyang")
silk = to_silk(audio_yunyang)
print(f"🔁 回发闭环: 人格语音 → {silk} (微信 #!SILK_V3, {os.path.getsize(silk)//1024}KB)")
print("\n✅ 全链路通过:语音输入 → 转写 → 人格回复 → 语音输出 → 微信格式")
def cmd_list():
print("可用人格:")
for k, p in PERSONAS.items():
print(f" {k:10s} {p['desc']}")
def main():
ap = argparse.ArgumentParser(description="voice-persona: 语音对话 + 人格音色")
sub = ap.add_subparsers(dest="cmd", required=True)
s1 = sub.add_parser("stt", help="语音文件转文字(支持微信 silk)")
s1.add_argument("file"); s1.add_argument("--model", default="small")
s1.add_argument("--language", default="zh")
s1.set_defaults(func=lambda a: print(stt_file(a.file, a.model, a.language)))
s2 = sub.add_parser("speak", help="文字→人格口吻→语音 mp3")
s2.add_argument("text"); s2.add_argument("--persona", default="xiaoyi")
s2.add_argument("--out", default="")
s2.set_defaults(func=lambda a: print(speak(personalize(a.text, a.persona), a.persona, a.out)))
s3 = sub.add_parser("chat", help="文字→人格回复文本")
s3.add_argument("text"); s3.add_argument("--persona", default="xiaoyi")
s3.add_argument("--llm", action="store_true", help="用 LLM 增强(需 LLM_API_KEY)")
s3.set_defaults(func=lambda a: print(personalize(a.text, a.persona, a.llm)))
s4 = sub.add_parser("demo", help="免 key 全链路自测")
s4.set_defaults(func=lambda a: cmd_demo())
s5 = sub.add_parser("list", help="列出人格")
s5.set_defaults(func=lambda a: cmd_list())
s6 = sub.add_parser("to_silk", help="音频→微信 silk 语音(回发语音气泡)")
s6.add_argument("audio"); s6.add_argument("--out", default="")
s6.set_defaults(func=lambda a: print(to_silk(a.audio, a.out)))
a = ap.parse_args()
a.func(a)
if __name__ == "__main__":
main()
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |