Skip to content

RECORDING.md

Latest commit

 

History

History
69 lines (50 loc) · 1.91 KB

RECORDING.md

File metadata and controls

69 lines (50 loc) · 1.91 KB

Custom Voice Recording Guide

Clone any voice—your own, a colleague, a character, or any speaker you like.

Requirements

Requirement Specification
Duration 10-20 seconds of speech
Format WAV
Sample rate 24kHz or higher
Channels Mono
Content Single speaker, no background noise
Speech Clear articulation, natural pace

Recording Methods

Option A: Record with sox

# Install sox if needed
brew install sox

# Record (speak naturally, press Ctrl+C when done)
rec -r 24000 -c 1 voice_ref.wav trim 0 20

Option B: Convert existing audio

# From Voice Memos, podcast clip, or any audio file
ffmpeg -i input.m4a -ar 24000 -ac 1 -t 20 voice_ref.wav

Option C: Use QuickTime Player

  1. Open QuickTime Player
  2. File → New Audio Recording
  3. Record 10-20 seconds of speech
  4. Save and convert with ffmpeg: 
    ffmpeg -i recording.m4a -ar 24000 -ac 1 voice_ref.wav

Tips for Best Results

  • Environment: Record in a quiet room. Closets work great for dampening echo.
  • Content: Read a paragraph naturally, as if explaining something to a colleague.
  • Tone: Avoid whispering, shouting, or exaggerated emotions.
  • Quality: Use a decent microphone if available. Built-in Mac mic works but external is better.

Installing Your Voice

Replace the bundled voice with your recording:

cp /path/to/your/voice_ref.wav ~/.claude/plugins/marketplaces/claude-mlx-tts/assets/default_voice.wav

Model Options

If you want faster inference at the cost of quality, edit scripts/tts-notify.py:

# Options (size/quality tradeoff):
MLX_MODEL = "mlx-community/chatterbox-turbo-fp16"   # ~4GB, best quality (default)
MLX_MODEL = "mlx-community/chatterbox-turbo-8bit"   # ~1GB, great quality
MLX_MODEL = "mlx-community/chatterbox-turbo-4bit"   # ~500MB, acceptable quality