Qwen3 TTS 12Hz 1.7B VoiceDesign 4bit

Qwen3-TTS 12Hz 1.7B VoiceDesign 4-bit

This model was converted to MLX format from Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign using mlx-audio version 0.3.0.

Refer to the original model card for authoritative details about the model architecture, supported capabilities, intended use, limitations, and license.

This quantized MLX version is designed for local inference on Apple Silicon and can be used through mlx-audio or through OpenVox.

Model Details

Base model	Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign
Format	MLX 4-bit quantized weights
Runtime	mlx-audio
Desktop app	OpenVox
Primary capability	Text-to-speech with natural-language voice design
Recommended hardware	Apple Silicon
Processing	Local and offline after the model has been downloaded

Use with OpenVox

OpenVox provides a graphical desktop interface for running compatible text-to-speech models locally on Apple Silicon without requiring Python or command-line setup.

With OpenVox, this model can be used to:

• Generate speech from text
• Create voices using natural-language descriptions
• Describe characteristics such as age, gender, tone, accent, pacing, pitch, and speaking style
• Save and reuse designed voices
• Generate and export long-form audio
• Compare different voice descriptions and generation settings
• Process text locally after the model has been downloaded
• Access local voice-generation workflows through API and MCP integrations

Example voice descriptions include:

A warm, confident female narrator in her thirties, speaking clearly with a neutral British accent and a calm, professional tone.

An energetic young male voice with a conversational American accent, medium-fast pacing, and an enthusiastic delivery.

A deep, mature documentary narrator with slow pacing, precise pronunciation, and a serious cinematic tone.

OpenVox is a third-party application and is not affiliated with or endorsed by Qwen, Alibaba Cloud, or the mlx-audio maintainers.

Use with mlx-audio

Install or update mlx-audio:

pip install -U mlx-audio

CLI Example

Replace the model identifier below with the Hugging Face repository in which this conversion is hosted.

python -m mlx_audio.tts.generate \
  --model theoracleguy/Qwen3-TTS-12Hz-1.7B-VoiceDesign-4bit \
  --text "Hello, this is a test."

Python Example

from mlx_audio.tts.utils import load_model
from mlx_audio.tts.generate import generate_audio

model = load_model(
    "theoracleguy/Qwen3-TTS-12Hz-1.7B-VoiceDesign-4bit"
)

generate_audio(
    model=model,
    text="Hello, this is a test.",
    file_prefix="test_audio",
)

Voice Design

This model is intended for generating a voice from a natural-language instruction rather than cloning a speaker from reference audio.

A voice-design instruction can describe attributes such as:

• Perceived age
• Gender presentation
• Accent or dialect
• Speaking style
• Emotional tone
• Pitch
• Pacing
• Energy
• Vocal texture
• Intended use, such as narration, dialogue, education, or announcements

The exact argument used to pass a voice-design instruction may depend on the installed mlx-audio version. Refer to the current mlx-audio documentation and the upstream Qwen model card for the supported generation interface.

Intended Use

This model may be useful for:

• Narration
• Accessibility and read-aloud tools
• Audiobook prototyping
• Character voice development
• Educational content
• Product demonstrations
• Local speech-generation applications
• Research and experimentation with controllable TTS

Responsible Use

Users are responsible for ensuring that generated speech is used lawfully and ethically.

Do not use generated audio to impersonate real individuals, mislead listeners, facilitate fraud, or create deceptive content without appropriate consent and disclosure.

License

This repository is distributed under the Apache 2.0 license.

The base model remains subject to the licensing terms and usage conditions stated in the original Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign model card.