Qwen3-TTS Voice Cloning

Voice cloning using Qwen3-TTS-12Hz-1.7B-Base with combined reference samples.

Setup

uv init --python 3.12
uv add qwen-tts soundfile torch numpy

Usage

uv run python clone_voice.py

Generates speech using 2 combined reference samples for improved voice quality.

Performance: ~111s for 3 outputs on MPS (Apple Silicon)

Reference Format

Reference files follow the pattern: samples/ref_N (without extension)

• ref_1.wav + ref_1.txt
• ref_2.wav + ref_2.txt

Add or remove references by editing ref_paths list in clone_voice.py. References are concatenated with 1s silence gaps.

Output

Generated audio samples:

output_0

I'm not the pheasant plucker, I'm the pheasant plucker's mate. I'm only plucking pheasants 'cause the pheasant plucker's running late

output_0.mp4

---

output_1

Extremely accurate and stunningly beautiful bespoke printer profiles transform creative print making to an extraordinary extent.

output_1.mp4

---

output_2

Generating code from AI prompts can lead to verbose code, or duplication of existing code instead of using an abstraction. But there are times when this is perfectly acceptable, such as when building proof of concepts, or when topics like program efficiency are unimportant.

output_2.mp4

---

Realtime TTS

Command Line

For fast realtime generation without re-processing reference audio:

# 1. Create voice model once
uv run python create_voice_prompt.py

# 2. Use for realtime generation
uv run python realtime_tts.py

Web UI Demo

Launch the Gradio web interface:

uv run python demo_app.py

Then open http://localhost:8000 in your browser.

See docs/realtime.md for details.

Configuration

See docs/configuration.md for generation parameters.

Edit syn_texts in clone_voice.py to customize synthesis text.

Model

Location: ~/LLMs/Qwen3-TTS/Qwen3-TTS-12Hz-1.7B-Base

Device: Auto-detects MPS or CPU

References

• Qwen3-TTS Model
• Example Code