diff --git a/xhs_lsposed_a16_fix/skill_voice_synth/SKILL.md b/xhs_lsposed_a16_fix/skill_voice_synth/SKILL.md new file mode 100644 index 0000000..dcdc6d1 --- /dev/null +++ b/xhs_lsposed_a16_fix/skill_voice_synth/SKILL.md @@ -0,0 +1,79 @@ +--- +name: voice-synth-send +description: 文本转语音并作为微信语音消息发送(edge-tts 云合成/豆包可选 → SILK → 微信内部 uploadvoice cgi)。全程在 Pixel 手机 Termux 内运行,无需本地 TTS 模型。 +version: 1.0 +category: communication +--- + +# 声音合成 + 微信语音发送 + +把一段文本合成为语音,作为**微信语音消息**发给指定联系人。**全程在手机 Termux 内跑** +(Hermes 所在环境),无需 Mac、无需本地 TTS 模型。 + +``` +文本 → TTS(edge-tts 免费云 / 豆包可选)→ MP3 + → ffmpeg 16k mono PCM → silk-wasm SILK V3(腾讯 \x02#!SILK_V3 变种) + → stage 进微信数据目录 → am broadcast SEND_VOICE_CGI(内部 uploadvoice cgi) +``` +接收方收到**可播放的语音条**,发送方本地也显示**已发送**气泡。依赖上游 skill +[wechat-internal-send](../skill_wechat_send/SKILL.md) 的 `SEND_VOICE_CGI` 链路。 + +## When to Use +- 用户要「发一条语音」「用语音念给某人」「语音播报」 +- Agent 主动语音通知 / 播报 + +## 关键澄清:edge-tts 不是本地模型 +edge-tts 是**微软 Edge 的免费云 TTS**(Python 库,把文本发到微软服务器合成、流式取回), +**不需要本地模型、不需要 API key、不占 GPU,只需要联网**。所以「手机上没本地 TTS 模型」 +不是问题。Hermes 作为 Termux app 有正常网络(它本来就在调 Kimi API),edge-tts 直接可用。 + +## 前置(本机 Termux 已具备,2026-07-07 核实) +- `python3` + `edge-tts` 7.2.8、`ffmpeg`、`node` + `silk-wasm` + (`/data/data/com.termux/files/usr/lib/node_modules/silk-wasm`) +- root(`su`)——用于把 SILK 放进微信数据目录 + 发广播 +- wechathook LSPosed 模块已装、微信在跑(`SEND_VOICE_CGI` 由它处理) + +## 用法 +```bash +# 在 Termux 内(Hermes 环境)直接跑: +bash voice_synth.sh "<要说的话>" [voice] [engine] + +# 例:edge-tts 男声发给文件传输助手(先用 filehelper 测试) +bash voice_synth.sh filehelper "你好,这是一条语音测试。" + +# 指定音色(女声) +bash voice_synth.sh yibooo "下午三点开会,记得带材料。" zh-CN-XiaoxiaoNeural + +# 豆包引擎(需额度 + 环境变量 DOUBAO_KEY,可发带音乐/音效/多角色的场景音频) +DOUBAO_KEY=xxxx bash voice_synth.sh yibooo "<豆包 seed-audio 场景 prompt>" "" doubao +``` +- ``:rcontact.username(`filehelper` / `wxid_xxx` / `yibooo`=火星人)。先用 filehelper 测。 +- `[voice]`:edge-tts 音色,默认 `zh-CN-YunxiNeural`(男声);女声 `zh-CN-XiaoxiaoNeural`。 +- `[engine]`:`edge`(默认,免费无额度)| `doubao`(seed-audio-1.0,需 `DOUBAO_KEY` + 账号额度)。 + +## 引擎对比 +| 引擎 | 成本 | 特点 | +|---|---|---| +| **edge-tts**(默认)| 免费,无额度 | 单人 TTS,多语言多音色,联网即用;日常播报首选 | +| **豆包 seed-audio-1.0** | 按字数计费(火山账号额度)| 可生成整段带**背景音乐/音效/多角色**的场景音频;额度用完报 `45000030 not granted` | + +> ⚠️ 额度/授权:豆包报 `code 45000030 [resource_id=volc.service_type.10074] requested resource not granted` +> = 账号未开通「音频生成/SeedAudio」或**字数额度用完**。控制台开通/充值后同 key 即可用。 + +## 真机验证(2026-07-07) +- edge-tts(6.9s)+ 豆包 seed-audio(8.2s)均端到端发火星人(yibooo)成功,多分块(6000 字节/片) + 上传完整、可播放;发送方本地显示已发送。SILK 头 `\x02#!SILK_V3` 与微信自录字节级一致。 + +## 坑 +- **edge-tts 要联网**:Hermes 原生跑没问题;但 `adb→su→termux` 起的进程拿不到 per-UID DNS + (`http=000`),别用那种方式测网络部分——在 Termux 里直接跑或让 Hermes 调。 +- **SILK 必须 WeChat 进程可读**:脚本已 stage 进 `/data/data/com.tencent.mm/files/` + + chown 微信 uid + restorecon;`/sdcard` 会 EACCES。 +- **TMPDIR**:脚本强制 `$PREFIX/tmp`(root/adb 环境会把 TMPDIR 顶成不可写的 /data/local/tmp)。 +- **不要自己实现 SILK 编码**:用 silk-wasm(腾讯变种),pysilk 等未必字节兼容。 + +## 相关 +- `voice_synth.sh` — 本 skill 脚本 +- `../skill_wechat_send/SKILL.md` — 底层 SEND_VOICE_CGI 发送链 +- `../WECHAT_SEND_API.md` §2 — 语音全合成逆向(uploadvoice cgi + 派发) +- memory `wechat-voice-send-internals.md` diff --git a/xhs_lsposed_a16_fix/skill_voice_synth/voice_synth.sh b/xhs_lsposed_a16_fix/skill_voice_synth/voice_synth.sh new file mode 100755 index 0000000..b4b2b23 --- /dev/null +++ b/xhs_lsposed_a16_fix/skill_voice_synth/voice_synth.sh @@ -0,0 +1,62 @@ +#!/data/data/com.termux/files/usr/bin/bash +# voice_synth.sh — synthesize speech and send it as a WeChat voice message. +# Runs FULLY ON-DEVICE in Termux (where Hermes lives): no Mac, no local TTS model. +# +# text → TTS (edge-tts cloud, free / or Doubao) → MP3 +# → ffmpeg 16k mono PCM → silk-wasm SILK V3 (Tencent \x02#!SILK_V3 variant) +# → stage into WeChat's data dir (uid + SELinux ctx) +# → am broadcast SEND_VOICE_CGI (internal uploadvoice cgi, our synthetic voice send) +# +# The recipient gets a playable voice AND the sender sees a sent bubble. +# +# Usage: +# voice_synth.sh "" [voice] [engine] +# talker id as in rcontact.username (e.g. filehelper, wxid_xxx, yibooo) +# what to say (UTF-8; for Doubao you can embed the rich scene prompt) +# [voice] edge-tts voice name (default zh-CN-YunxiNeural male; e.g. zh-CN-XiaoxiaoNeural female) +# [engine] edge (default, free cloud, no quota) | doubao (needs DOUBAO_KEY + quota) +# +# Test with filehelper first. edge-tts needs internet (it's a cloud service, not a local model). +set -e + +WXID="$1"; TEXT="$2"; VOICE="${3:-zh-CN-YunxiNeural}"; ENGINE="${4:-edge}" +[ -z "$WXID" ] || [ -z "$TEXT" ] && { echo "usage: voice_synth.sh \"\" [voice] [engine]"; exit 1; } + +PREFIX=/data/data/com.termux/files/usr +SILK_WASM="$PREFIX/lib/node_modules/silk-wasm" +export TMPDIR="$PREFIX/tmp"; mkdir -p "$TMPDIR" # Termux tmp (root/adb env may force /data/local/tmp) +WORK=$(mktemp -d "$TMPDIR/voicesynth.XXXXXX") +trap 'rm -rf "$WORK"' EXIT +MP3="$WORK/tts.mp3"; PCM="$WORK/tts.pcm"; SILK="$WORK/tts.silk" + +# ── 1. text → MP3 ──────────────────────────────────────────────────────────── +case "$ENGINE" in + edge) + echo "[1/3] edge-tts ($VOICE) → MP3" + edge-tts --voice "$VOICE" --text "$TEXT" --write-media "$MP3" >/dev/null + ;; + doubao) + : "${DOUBAO_KEY:?doubao engine needs DOUBAO_KEY env var}" + echo "[1/3] Doubao seed-audio-1.0 → MP3" + PROMPT=$(python3 -c "import json,sys;print(json.dumps(sys.argv[1]))" "$TEXT") + curl -s --max-time 120 -X POST 'https://openspeech.bytedance.com/api/v3/tts/create' \ + -H 'Content-Type: application/json' -H "X-Api-Key: $DOUBAO_KEY" \ + --data-raw "{\"model\":\"seed-audio-1.0\",\"text_prompt\":$PROMPT,\"audio_config\":{\"format\":\"mp3\",\"sample_rate\":48000},\"watermark\":{}}" \ + | python3 -c "import json,sys,base64; d=json.load(sys.stdin); sys.exit('Doubao error: '+str(d)) if 'audio' not in d else open('$MP3','wb').write(base64.b64decode(d['audio']))" + ;; + *) echo "unknown engine: $ENGINE (use edge|doubao)"; exit 1 ;; +esac + +# ── 2. MP3 → 16k mono PCM → SILK V3 ────────────────────────────────────────── +echo "[2/3] MP3 → 16k mono PCM → SILK V3" +ffmpeg -y -i "$MP3" -ar 16000 -ac 1 -f s16le "$PCM" -loglevel error +DUR=$(node -e "const s=require('$SILK_WASM'),fs=require('fs');s.encode(fs.readFileSync('$PCM'),16000).then(r=>{fs.writeFileSync('$SILK',r.data);process.stdout.write(String(Math.round(r.duration)));})") +[ -s "$SILK" ] || { echo "SILK encode failed"; exit 1; } + +# ── 3. stage into WeChat data dir + fire SEND_VOICE_CGI (root) ──────────────── +echo "[3/3] stage SILK + SEND_VOICE_CGI → $WXID (${DUR}ms)" +WX_UID=$(su -c "stat -c %U /data/user/0/com.tencent.mm" | tr -d '\r') +STAGE=/data/data/com.tencent.mm/files/hook_voice_cgi.silk +su -c "cp '$SILK' '$STAGE' && chown $WX_UID:$WX_UID '$STAGE' && chmod 600 '$STAGE'; restorecon '$STAGE' 2>/dev/null; \ + am broadcast -a com.hermes.wechat.SEND_VOICE_CGI --es to '$WXID' --es src_path '$STAGE' --ei duration $DUR" +echo "done: sent ${DUR}ms voice to $WXID via $ENGINE."