Run Qwen3-TTS text-to-speech models locally on your Mac with Apple Silicon.

I've made another repo using MLX (Apple's machine learning framework) that runs faster and uses less memory. If you want the best Mac experience, use that instead:

qwen3-tts-apple-silicon - Native MLX version, better performance

This repo uses PyTorch with MPS backend. Still works great, but the MLX version is more optimized for Mac.

• Text-to-speech with multiple voice options
• Voice cloning from audio samples
• Custom voice design with text prompts
• Optimized for M1/M2/M3/M4 chips
• Simple terminal interface

========================================
 Qwen3-TTS Studio (Mac Optimized)
========================================
 High Quality (1.7B)
  1. Custom Voice
  2. Voice Design
  3. Voice Clone

 Fast / Low Res (0.6B)
  4. Custom Voice
  5. Voice Clone

  6. Exit

Select Option: 1

[Loader] Loading CustomVoice (Pro)...
[System] M3/M4 detected. Using bfloat16 for efficiency.
[Loader] Model loaded successfully.

--- CustomVoice (Pro) Mode ---

Available Speakers:
  [English]: Ryan, Aiden
  [Chinese]: Vivian, Serena, Uncle_Fu, Dylan, Eric
  [Japanese]: Ono_Anna
  [Korean]: Sohee

Select Speaker: Ryan
[Selection] Speaker set to: Ryan

Enter text (or 'exit'): Hello, this is a test of the text to speech system.
[Status] Generating...
[IO] Saved to: outputs/CustomVoice_Pro/14-23-45_Hello_this_is_a_tes.wav

brew install llvm@20

Add to your shell (put in ~/.zshrc):

export PATH="/usr/local/opt/llvm@20/bin:$PATH"
export CMAKE_PREFIX_PATH="/usr/local/opt/llvm@20"

git clone https://github.com/kapi2800/qwen3-tts-mac.git
cd qwen3-tts-mac

git clone https://github.com/QwenLM/Qwen3-TTS.git

uv init -p 3.12
uv sync
uv add ./Qwen3-TTS
uv add 'huggingface-hub[cli]'
uv add 'numpy<2'

Pick one based on your Mac's RAM:

8GB RAM - Use Mini (0.6B)

huggingface-cli download Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice --local-dir ./models/Qwen3-TTS-12Hz-0.6B-CustomVoice

16GB+ RAM - Use Pro (1.7B)

huggingface-cli download Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --local-dir ./models/Qwen3-TTS-12Hz-1.7B-CustomVoice

uv run python main.py

Edit these values at the top of main.py:

MAX_CHAR_LIMIT = 600                  # Max text length per generation
THERMAL_COOLDOWN_SECONDS = 0.3        # Pause between generations (reduce heat)
ENABLE_TORCH_COMPILE = False          # Experimental speedup

• First generation is always slower (model warmup)
• Keep text under 600 characters
• Use Mini models if your Mac runs hot
• Audio files save to outputs/ folder

Latest:

• Dynamic token optimization for faster short text
• Memory-efficient model loading
• Thermal management between generations
• Auto bfloat16 for M3/M4 chips

If you made it this far and this actually worked for you, maybe hit that star button. Or don't. I'm just a readme.