blue-onnx 0.1.0

Hebrew Text-to-Speech inference using ONNX and TensorRT

Blue

Text-to-Speech inference using ONNX Runtime with optional TensorRT acceleration.

✨ Demo

 

🎙️ Human-sounding TTS in Hebrew, English, Spanish, Italian & German — try samples and the live demo on the site.

Installation

Install the core dependencies:

uv sync

For CUDA (GPU) support:

uv sync --extra gpu

Download Models

uv run hf download notmax123/blue-onnx --repo-type model --local-dir ./onnx_models \
  --exclude "voices/all_voices/**"

Optional:

• Hebrew G2P:

  wget -O model.onnx https://huggingface.co/thewh1teagle/renikud/resolve/main/model.onnx
  

• 2000+ voice JSONs:

  uv run hf download notmax123/blue-onnx voices/all_voices/ --repo-type model --local-dir ./onnx_models
  

• PyTorch weights (export new voices): uv sync --extra export then

  uv run hf download notmax123/blue --repo-type model --local-dir ./pt_models
  

Usage

Examples use voices/female1.json from this repo. After the optional voice download, use paths under onnx_models/voices/all_voices/ (manifest.tsv lists them).

Here is a basic example of how to use BlueTTS in Python:

import soundfile as sf
from src.blue_onnx import BlueTTS

tts = BlueTTS(
    onnx_dir="onnx_models",
    style_json="voices/female1.json",
    renikud_path="model.onnx",
)

# Single language
samples, sr = tts.synthesize("שלום, זהו מודל דיבור בעברית.", lang="he")
sf.write("output.wav", samples, sr)

# Mixed languages
mixed = "שלום לכולם, <en>welcome to the presentation</en>, <es>espero que lo disfruten</es>."
samples, sr = tts.synthesize(mixed, lang="he")
sf.write("mixed_output.wav", samples, sr)

Running Examples

You can run the provided example scripts to test the model. Outputs will be saved in the examples/out/ directory.

# Generate samples for individual languages
uv run python examples/hebrew.py
uv run python examples/english.py
uv run python examples/spanish.py
uv run python examples/italian.py
uv run python examples/german.py

# Generate a mixed-language sample
uv run python examples/mixed.py

# Run the CLI app
uv run python examples/app.py --lang en --text "Hello world."

TensorRT (NVIDIA GPUs Only)

For faster inference on NVIDIA GPUs, you can build TensorRT engines.

1. Install TensorRT dependencies:

uv sync --extra tensorrt
uv pip install tensorrt-cu12  # installed separately due to astral-sh/uv#14313

2. Build the engines (see scripts/README.md for details):

uv run python scripts/create_tensorrt.py \
  --onnx_dir onnx_models --engine_dir trt_engines --precision fp16 --config config/tts.json

(Note: The examples/all_langs_and_mix.py --tensorrt flag is currently bugged).

Papers

@ARTICLE{2025arXiv250323108K,
       author = {{Kim}, Hyeongju and {Yang}, Jinhyeok and {Yu}, Yechan and {Ji}, Seunghun and {Morton}, Jacob and {Bous}, Frederik and {Byun}, Joon and {Lee}, Juheon},
        title = "{SupertonicTTS: Towards Highly Efficient and Streamlined Text-to-Speech System}",
      journal = {arXiv e-prints},
     keywords = {Audio and Speech Processing, Machine Learning, Sound},
        pages = {arXiv:2503.23108},
}
@article{kim2025training,
  title={Training Flow Matching Models with Reliable Labels via Self-Purification},
  author={Kim, Hyeongju and Yu, Yechan and Yi, June Young and Lee, Juheon},
  journal={arXiv preprint arXiv:2509.19091},
  year={2025}
}
@misc{yi2025robustttstrainingselfpurifying,
      title={Robust TTS Training via Self-Purifying Flow Matching for the WildSpoof 2026 TTS Track}, 
      author={June Young Yi and Hyeongju Kim and Juheon Lee},
      year={2025},
      eprint={2512.17293},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2512.17293}, 
}

Acknowledgments

This project uses renikud for Hebrew G2P. Special thanks to thewh1teagle for his work on Hebrew phonemization.

License

MIT