DraculaFlow-TTS-mlx-8bit

This repository contains the 8-bit quantized MLX-converted weights for DraculaFlow-TTS, optimized for fast and memory-efficient inference on Apple Silicon using mlx-audio.

• Original Hugging Face Model: LunaFox/DraculaFlow-TTS

---

Installation

Install mlx-audio:

pip install mlx-audio

---

Usage

1. Command Line Interface (CLI)

Generate audio clips directly from the terminal using mlx_audio.tts.generate:

python -m mlx_audio.tts.generate \
  --model LunaFox/DraculaFlow-TTS-mlx-8bit \
  --text "We are smoking joint out of a custom carved bong." \
  --voice dracula_flow \
  --output_path ./output.wav

To play the audio immediately upon generation, add the --play flag:

python -m mlx_audio.tts.generate \
  --model LunaFox/DraculaFlow-TTS-mlx-8bit \
  --text "This shit ain't nothing to me man." \
  --voice dracula_flow \
  --play

2. Python API

You can also generate audio clips within a Python script:

from mlx_audio.tts.generate import generate_audio

generate_audio(
    text="We are smoking joint out of a custom carved bong.",
    model="LunaFox/DraculaFlow-TTS-mlx-8bit",
    voice="dracula_flow",
    output_path="./output.wav",
)

---

Details

• Architecture: Qwen3 TTS (qwen3_tts)
• Quantization: 8-bit (group size: 64)
• Speaker Voice ID: dracula_flow
• Framework: MLX / mlx-audio