메뉴
Text-to-speech and speech-to-text on Apple Silicon using Vox CLI (Qwen3-TTS + MLX)
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Decide whether a task in the current project is worth running as an agent loop. Analyzes the repo for evidence first (tests/CI/bench scripts = available verifiers; issue+PR queues = recurring work; module boundaries), then interviews the user one question at a time on the genuine decisions, and returns one of three verdicts — don't loop (stay in the loop yourself), timer loop (/loop, /schedule), or goal loop (/goal) — with cited evidence and, when looping is warranted, a drafted four-part contract (goal / verification / boundary / stop) bound to real commands found in the repo. Recommending NO loop is a first-class outcome. Use when the user asks 值不值得 loop / should I loop this / 要不要上 /goal / 这个活能不能挂个循环自动跑, or wants to apply loop engineering to a project.
GitHub backlog governance manager-loop. Triage open issues (type + routing labels), complete thin descriptions, maintain a bounded ready queue (Todo ≤ 5) on a GitHub Projects board, and repair board drift (closed issue still "In Progress" etc.). Config-driven — reads .claude/backlog-manager.yaml from the target repo; runs an init flow to generate it if missing. DRY-RUN by default, pass "apply" to execute writes. Use for recurring backlog grooming / issue triage of any GitHub repo, standalone or driven by /loop. Requires gh CLI with repo + project scopes.
Generate images (Nano Banana Pro / Imagen) and videos (Veo 3.1) through Google Flow using the account's Ultra/Pro SUBSCRIPTION credits instead of the metered Gemini/Vertex API — no per-call API cost. Use when the user wants to create a thumbnail, cover, poster, b-roll clip, image, or short video with AI and wants to avoid API billing, or explicitly mentions Flow / gflow / Veo / Nano Banana via subscription. Wraps the gflow-cli tool; encodes this machine's Flow new-UI quirk (PREFER_CLASSIC) and credit-safe retry rules learned the hard way. NOT for the paid Gemini-API nano-banana-pro path (that costs money) — this is the subscription path.
Generate a printer-receipt styled PNG bill of AI token usage and cost from local ccusage data, for the 绿皮火车 channel. Use when the user wants a token usage "账单"/"小票"/"收银台"/receipt/invoice, a shareable spend breakdown by model / input / output / cached / by day, or 节目素材 about token 消耗/花费. Triggers on "生成账单", "token 收据", "做张小票", "用量账单", "token receipt", "spend breakdown image".
本地把音频/视频文件或在线视频 URL 转录成文字稿(txt/srt/vtt/json)。基于 Apple Silicon 上的 mlx_whisper,支持中英文等多语言、自动语言检测、模型规格选择。当用户想要"转录""生成字幕/文稿/transcript""把这段音频/视频转成文字""提取台词""做 SRT 字幕",或给出一个音视频文件/YouTube 等链接要文字内容时使用。
Jordan Peterson(乔丹·彼得森,「龙虾教授」)的思维框架与表达方式。基于著作、长访谈/辩论 transcript、X发帖、外部批评、决策记录、完整时间线共6维度216个来源(一手占比约48%)的深度调研, 提炼6个核心心智模型、8条决策启发式和完整的表达DNA,并内置其已被外部验证的失效模式标注。 用途:作为思维顾问,用彼得森的视角分析个人困境、意义危机、责任与成长、文化争议、叙事与神话。 当用户提到「用彼得森的视角」「Jordan Peterson会怎么看」「彼得森」「皮特森」「龙虾教授」「JP模式」「JBP」 「peterson perspective」「人生十二法则」「12条法则」时使用。 即使用户只是说「整理好你的房间」「混乱与秩序」「先承担责任」 「帮我用彼得森的角度想想」「切换到彼得森」也应触发。 用户讨论「躺平」「内卷」「精神内耗」「年轻人迷茫」「人生没有意义」等话题且明确想要某种视角分析或建议时也可触发。 不要在用户只是泛泛倾诉、求安慰、查心理学概念,或寻求真实医疗/心理咨询时触发—— 这是思维框架skill,不是心理治疗。
| name | tts |
| description | Text-to-speech and speech-to-text on Apple Silicon using Vox CLI (Qwen3-TTS + MLX) |
Local text-to-speech, speech-to-text, and voice cloning powered by Qwen3-TTS/ASR + MLX on Apple Silicon.
A powerful local TTS/STT skill based on Vox CLI that runs entirely on Apple Silicon Macs. Features speech synthesis with preset and custom voices, voice cloning from audio samples, speech recognition with subtitle generation, and batch processing. All models run locally via MLX - your data never leaves your machine.
Use this skill when users:
python scripts/vox_tts.py speak "Hello, this is a test" [OPTIONS]
python scripts/vox_tts.py transcribe audio.wav [OPTIONS]
python scripts/vox_tts.py clone "Text to speak" --ref voice_sample.wav [OPTIONS]
Speak text with default voice:
python scripts/vox_tts.py speak "Hello world" -o ./output
Speak text and play immediately:
python scripts/vox_tts.py speak "Hello world" --play
Speak with a specific voice:
python scripts/vox_tts.py speak "Hello world" --voice Chelsie -o ./output
Speak with emotion/style instruction:
python scripts/vox_tts.py speak "I can't believe it!" --instruct "excited and surprised" -o ./output
Speak Chinese text:
python scripts/vox_tts.py speak "你好，世界" --voice Chelsie -o ./output
Clone a voice and speak:
python scripts/vox_tts.py clone "Text in the cloned voice" --ref sample.wav -o ./output
Register a cloned voice for reuse:
python scripts/vox_tts.py clone --ref sample.wav --register my-voice
python scripts/vox_tts.py speak "Now using my custom voice" --voice my-voice
Design a voice from description:
python scripts/vox_tts.py design "Hello everyone" --desc "A warm, friendly female voice with a slight British accent" -o ./output
Transcribe audio to text:
python scripts/vox_tts.py transcribe recording.wav
Transcribe with subtitles:
python scripts/vox_tts.py transcribe recording.wav --subtitle srt -o ./output
Batch TTS from file (one line per utterance):
python scripts/vox_tts.py batch texts.txt --voice Chelsie -o ./output
Use large model for higher quality:
python scripts/vox_tts.py speak "High quality speech" --model large -o ./output
List available voices:
python scripts/vox_tts.py voices
speak - Text-to-Speech| Argument | Description | Default |
|---|---|---|
text | Text to synthesize (required) | - |
-o, --output | Output directory | Current directory |
-v, --voice | Voice name (see voices command) | Chelsie |
-m, --model | Model size: small, large, large-hq | small |
-s, --speed | Speech speed multiplier | 1.0 |
-i, --instruct | Emotion/style instruction | None |
--play | Play audio after generation | False |
--subtitle | Generate subtitle: srt or vtt | None |
transcribe - Speech-to-Text| Argument | Description | Default |
|---|---|---|
audio | Audio file path (required) | - |
-o, --output | Output directory | Current directory |
--subtitle | Output format: srt, vtt, json | plain text |
--language | Source language | auto-detect |
clone - Voice Cloning| Argument | Description | Default |
|---|---|---|
text | Text to synthesize | None |
--ref | Reference audio file (3+ seconds, required) | - |
--register | Register as reusable voice name | None |
-o, --output | Output directory | Current directory |
design - Voice Design| Argument | Description | Default |
|---|---|---|
text | Text to synthesize (required) | - |
--desc | Voice description in natural language (required) | - |
-o, --output | Output directory | Current directory |
batch - Batch Processing| Argument | Description | Default |
|---|---|---|
file | Text file, one utterance per line (required) | - |
-v, --voice | Voice name | Chelsie |
-o, --output | Output directory | Current directory |
# Requirements: Apple Silicon Mac (M1/M2/M3/M4), Python 3.10+, macOS 13+
# Install via pipx (recommended, global CLI)
brew install pipx
pipx ensurepath
git clone https://github.com/3Craft/tts.git /tmp/vox-tts
cd /tmp/vox-tts && pipx install .
# Or install via pip in a venv
pip install -e /path/to/tts
# Optional: Chinese text support
pipx inject vox-cli 'misaki[zh]'
# Optional: Japanese text support
pipx inject vox-cli 'misaki[ja]'
OutputDir/
├── output.wav # Generated audio
└── output.srt # Subtitle (if --subtitle srt)
OutputDir/
├── 001_first_line.wav
├── 002_second_line.wav
└── ...
OutputDir/
├── recording.txt # Plain text transcript
└── recording.srt # Subtitle (if --subtitle srt)
When user requests TTS, STT, or voice cloning:
Read skill documentation:
view("/mnt/skills/user/tts/SKILL.md")
Check if vox is installed:
python /mnt/skills/user/tts/scripts/vox_tts.py check
Install if needed:
brew install pipx && pipx ensurepath
git clone https://github.com/3Craft/tts.git /tmp/vox-tts
cd /tmp/vox-tts && pipx install .
Execute command:
# TTS
python /mnt/skills/user/tts/scripts/vox_tts.py speak "TEXT" \
--voice Chelsie -o /mnt/user-data/outputs
# STT
python /mnt/skills/user/tts/scripts/vox_tts.py transcribe audio.wav \
-o /mnt/user-data/outputs
# Voice cloning
python /mnt/skills/user/tts/scripts/vox_tts.py clone "TEXT" \
--ref sample.wav -o /mnt/user-data/outputs
Present files to user:
present_files(["/mnt/user-data/outputs/..."])
Run vox voices to see all preset voices. Some commonly used:
| Voice | Description |
|---|---|
| Chelsie | Default female voice |
| Ethan | Male voice |
Custom voices can be registered via clone --register.
large-hq for highest qualityvox serve) keeps models in memory for sub-second responseQ: First run is slow?
A: Models are downloaded on first use (1-3 GB). Subsequent runs are much faster. Use daemon mode (vox serve) for instant response.
Q: "Not Apple Silicon" error? A: Vox CLI requires Apple Silicon (M1/M2/M3/M4). It uses MLX which only runs on Apple's Neural Engine.
Q: Chinese/Japanese text sounds wrong?
A: Install language support: pipx inject vox-cli 'misaki[zh]' or pipx inject vox-cli 'misaki[ja]'
Q: Out of memory?
A: Use the small model (default) instead of large. Close other apps to free memory.
Q: How to use a custom voice persistently?
A: Register it: vox clone --ref sample.wav --register my-voice, then use --voice my-voice.
User: "帮我把这段文字转成语音：今天天气真好"
Claude:
python /mnt/skills/user/tts/scripts/vox_tts.py speak "今天天气真好" \
--voice Chelsie -o /mnt/user-data/outputs --play
User: "Transcribe this audio file to SRT subtitles"
Claude:
python /mnt/skills/user/tts/scripts/vox_tts.py transcribe recording.wav \
--subtitle srt -o /mnt/user-data/outputs
User: "用这段录音克隆一个声音，然后用它朗读一段话"
Claude:
# Register the cloned voice
python /mnt/skills/user/tts/scripts/vox_tts.py clone \
--ref sample.wav --register custom-voice
# Speak with the cloned voice
python /mnt/skills/user/tts/scripts/vox_tts.py speak "这是用克隆声音朗读的文字" \
--voice custom-voice -o /mnt/user-data/outputs --play
v1.0 (Current)