Vox TTS / STT (Apple Silicon)

Local text-to-speech, speech-to-text, and voice cloning powered by Qwen3-TTS/ASR + MLX on Apple Silicon.

Description

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.

When to Use

Use this skill when users:

• Want to convert text to speech, "read this aloud", "generate audio from text"
• Ask for "TTS", "text to speech", "语音合成", "文字转语音", "朗读"
• Want to transcribe audio to text, "STT", "speech to text", "语音识别", "转录"
• Need voice cloning from a sample, "clone this voice", "克隆声音"
• Want to generate subtitles (SRT/VTT) from audio
• Ask for batch text-to-speech conversion
• Mention "vox" or "Qwen3-TTS"

Features

• Text-to-Speech: Synthesize speech with 20+ preset voices
• Speech-to-Text: Transcribe audio with word-level timestamps
• Voice Cloning: Clone a voice from a 3+ second audio sample
• Voice Design: Describe a voice in natural language and generate it
• Subtitle Generation: Output SRT/VTT with word-level timestamps
• Batch Processing: Process multiple texts in one model-load cycle
• Daemon Mode: Keep models resident in memory for faster subsequent calls
• Multiple Models: 0.6B (small/fast) and 1.7B (large/quality) parameter models
• 100% Local: All processing runs on-device via MLX

Usage

Basic Syntax

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]

Common Scenarios

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

Arguments

speak - Text-to-Speech

transcribe - Speech-to-Text

clone - Voice Cloning

design - Voice Design

batch - Batch Processing

Dependencies

# 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]'

Output Structure

Single TTS

OutputDir/
├── output.wav              # Generated audio
└── output.srt              # Subtitle (if --subtitle srt)

Batch TTS

OutputDir/
├── 001_first_line.wav
├── 002_second_line.wav
└── ...

Transcription

OutputDir/
├── recording.txt           # Plain text transcript
└── recording.srt           # Subtitle (if --subtitle srt)

Claude Integration

When user requests TTS, STT, or voice cloning:

1. Read skill documentation:

   view("/mnt/skills/user/tts/SKILL.md")

2. Check if vox is installed:

   python /mnt/skills/user/tts/scripts/vox_tts.py check

3. 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 .

4. 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

5. Present files to user:

   present_files(["/mnt/user-data/outputs/..."])

Available Voices

Run vox voices to see all preset voices. Some commonly used:

Custom voices can be registered via clone --register.

How It Works

TTS Workflow

1. Text Processing: Tokenize and normalize input text
2. Model Inference: Run Qwen3-TTS model via MLX on Apple Silicon GPU
3. Audio Generation: Generate WAV audio at 24kHz
4. Post-Processing: Apply speed adjustment, generate subtitles if requested

STT Workflow

1. Audio Loading: Read audio file and resample if needed
2. Model Inference: Run Qwen3-ASR model via MLX
3. Decoding: Generate text with word-level timestamps
4. Output: Format as plain text, SRT, VTT, or JSON

Under the Hood

• Models are downloaded from HuggingFace on first use (~1-3 GB)
• 8-bit quantization by default for memory efficiency
• bf16 mode available via large-hq for highest quality
• Daemon mode (vox serve) keeps models in memory for sub-second response

Common Issues

Q: 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.

Example Conversations

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

Limitations

• Apple Silicon only: Requires M1/M2/M3/M4 Mac (MLX dependency)
• macOS 13+: Minimum OS version requirement
• Memory: Large model needs ~4GB RAM; small model ~2GB
• Languages: Best quality for English and Chinese; other languages may vary
• Voice cloning: Requires 3+ seconds of clear speech reference audio

Version History

v1.0 (Current)

• Initial release wrapping Vox CLI v0.3.3
• TTS, STT, voice cloning, voice design
• Batch processing and subtitle generation
• Daemon mode support

References

• Vox CLI GitHub
• MLX Audio
• Qwen3-TTS