Dia is a 1.6B parameter text-to-speech model developed by Nari Labs, designed to generate highly realistic dialogues directly from transcripts. It supports emotional conditioning and non-verbal cues such as laughter, coughing, and more.

🧪 Pretrained model checkpoints and inference code are publicly available to accelerate research and experimentation.
💬 English only (for now).
🔬 Demo available on Hugging Face Spaces.

This is a community-enhanced fork of the original nari-labs/dia repository. It focuses on performance optimizations and accessibility for personal hardware.

• 🧠 Quantization (4-bit & INT8)
  Reduces VRAM usage by nearly 50%. The model now runs on GPUs with just 8GB VRAM, consuming less than 6GB in practice.

• ⚡ Flash Attention Integration
  Experimental support for Flash Attention to speed up inference and reduce memory footprint.

• 🔄 Continued Development
  Ongoing efforts to improve inference speed, memory efficiency, and model accessibility.

• 🔈 Generate expressive dialogue using [S1], [S2] tags.
• 🤖 Realistic non-verbal sounds: (laughs), (sighs), (coughs), etc.
• 🧬 Optional voice cloning via reference audio.
• 🪄 Supports speaker conditioning and output diversity.
• 💡 Simple Python API for generation and saving audio.

pip install git+https://github.com/rzafiamy/dia-q4-bit.git

git clone https://github.com/rzafiamy/dia-q4-bit.git
cd dia
uv run app.py

Or using a Python virtual environment:

git clone https://github.com/rzafiamy/dia-q4-bit.git
cd dia
python -m venv .venv
source .venv/bin/activate
pip install -e .
python app.py

ℹ️ Voices vary between runs unless you condition output with an audio prompt or set a fixed seed.

from dia.model import Dia

def get_model(quantize=False, quantize_4bit=False):
    print("[INFO] Loading Dia model...")
    return Dia.from_pretrained(
        "nari-labs/Dia-1.6B",
        compute_dtype="float16",
        quantize=quantize,
        quantize_4bit=quantize_4bit
    )

model = get_model(quantize_4bit=True)  # or quantize=True for int8

text = "[S1] Dia is fast and memory-efficient! [S2] And it sounds great. (laughs)"

output = model.generate(text, use_torch_compile=True, verbose=True)
model.save_audio("simple.mp3", output)

• quantize=True: Use INT8 quantization
• quantize_4bit=True: Use 4-bit quantization
• Both reduce VRAM usage significantly (under 6 GB), perfect for 8 GB consumer GPUs.
• Only one of these options should be enabled at a time.

📢 CLI and PyPI package coming soon.

Tested on PyTorch 2.0+ and CUDA 12.6. CPU support coming soon.

Apache License 2.0 – see the LICENSE file for full details.

This model is for research and educational use only. By using it, you agree not to:

• Mimic real identities without consent.
• Generate misleading or harmful content.
• Use it for illegal or malicious purposes.

• 4-bit / INT8 quantization support.
• Flash Attention integration.
• CPU inference support.
• Docker support (incl. ARM/MacOS).
• Public PyPI release & CLI.
• Larger model versions.

This fork is maintained by the open-source community.
PRs are welcome! Join us on Discord to collaborate or share ideas.

• Nari Labs for the original Dia model.
• Google TPU Research Cloud for computing resources.
• Research inspiration from: SoundStorm, Parakeet, and Descript Audio Codec.
• Hugging Face for hosting weights and demo space.