Blue

Text-to-speech inference with ONNX Runtime; optional TensorRT acceleration on NVIDIA GPUs.

   

Hebrew, English, Spanish, Italian, and German — samples and a live demo on the site and Space above.

Install

Requires Python 3.12+ (see requires-python in pyproject.toml).

Users (PyPI)

The PyPI package is inference only (ONNX TTS in Python). It does not ship training code or the repo examples.

pip install "blue-onnx"

For Rust ONNX inference, see blue-rs.

Inference with pip in three steps: (1) install as above, (2) put ONNX in ./onnx_models (see Models — hf download from blue-onnx-v2, or the int8 INT8 bundle into another folder), (3) copy a style JSON (e.g. from voices/) and Hebrew G2P model.onnx (see Models), then use this from your project:

import soundfile as sf
from blue_onnx import BlueTTS

tts = BlueTTS(
    onnx_dir="onnx_models",
    style_json="voices/female1.json",
    renikud_path="model.onnx",
)
s, sr = tts.synthesize("Hello", lang="en")
sf.write("out.wav", s, sr)

Optional accelerators (PyPI): install the extra, then drop the stock CPU wheel so a single build owns onnxruntime:

• Intel OpenVINO: pip install "blue-onnx[openvino]" then pip uninstall onnxruntime
• NVIDIA CUDA: pip install "blue-onnx[gpu]" then pip uninstall onnxruntime

This repository

git clone https://github.com/maxmelichov/BlueTTS.git
cd BlueTTS
uv sync

Optional extras:

uv sync --extra openvino   # Intel OpenVINO EP (then: uv pip uninstall onnxruntime)
uv sync --extra gpu        # NVIDIA CUDA ORT (then: uv pip uninstall onnxruntime)

The default environment uses the stock onnxruntime CPU wheel. OpenVINO and CUDA (gpu) are optional; each adds a second ONNX Runtime distribution until you remove the stock CPU wheel so the accelerator build owns the onnxruntime import (uv pip uninstall onnxruntime, or the same with pip after a PyPI install with [openvino] or [gpu]). Do not combine the openvino and gpu extras. For TensorRT, use uv sync --extra tensorrt in TensorRT. For voice or ONNX export, add --extra export in Models (with the PyTorch checkpoint download).

Models

FP32 bundle (recommended) — notmax123/blue-onnx-v2: onnxslim–cleaned, full precision. Includes the four core graphs (text_encoder, vector_estimator, vocoder, duration_predictor), runtime tts.json / vocab.json, and ONNX graphs for zero-shot voice conversion from a reference clip (codec_encoder, style_encoder, duration_style_encoder).

uv run hf download notmax123/blue-onnx-v2 --repo-type model --local-dir ./onnx_models

INT8 bundle (experimental) — notmax123/bluev2-onnx-int8: weight-only quantized graphs for a smaller footprint; not slimmed after export. Same filenames and layout as the FP32 bundle—set onnx_dir to the downloaded folder. Prefer FP32 for best quality.

uv run hf download notmax123/bluev2-onnx-int8 --repo-type model --local-dir ./onnx_int8

Local export with exports/export_onnx.py: use --slim for FP32 (matches the published FP32 style) or --int8 without --slim for INT8; INT8 remains experimental.

The Hub bundles do not include per-voice style JSON; use the sample voices/*.json from this repository (or on GitHub), or export a new voice from a reference clip (see exports/README.md, PyTorch weights below). If you use pip without uv, the same CLI is available after install because blue-onnx depends on huggingface-hub — run hf download ... with the same arguments and point style_json at a file under voices/ (e.g. voices/female1.json).

Optional

• Hebrew G2P:

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

• PyTorch checkpoints (notmax123/blue-v2) for exporting new voice JSON and ONNX: uv sync --extra export then

  uv run hf download notmax123/blue-v2 --repo-type model --local-dir ./pt_models

Usage

Examples below use voices/female1.json from this repo, or a JSON you produced with exports/export_new_voice.py.

Quick start

After the BlueTTS(...) setup in Install (PyPI) or the same paths in a clone (import soundfile as sf and build tts as there):

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

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)

If you are editing this repo without installing the package, use from src.blue_onnx import BlueTTS (as in examples/) or put src on PYTHONPATH.

Examples

Get models into ./onnx_models (see Models) and voices/*.json from the repo, then from the repository root use either uv or a PyPI install:

uv run python examples/basic.py   # he / en / es / it / de + mixed in one run
uv run python examples/mixed.py
uv run python examples/app.py --lang en --text "Hello world."

With pip install "blue-onnx" (and the same onnx_models + voices/), the same files use import blue_onnx automatically; a dev tree without the package falls back to src.blue_onnx. If your graphs live somewhere else, set ONNX_DIR for basic.py / mixed.py, or pass --onnx-dir to app.py. Default app output: examples/out/app_output.wav.

Edit voice JSON paths or ONNX_DIR if your layout differs. See examples/voices.md for app.py and voice selection.

TensorRT (NVIDIA only)

1. Dependencies:

uv sync --extra tensorrt
uv pip install tensorrt-cu12   # separate install; see astral-sh/uv#14313

2. Build engines (see also exports/README.md):

uv run python exports/create_tensorrt.py \
  --onnx_dir onnx_models --engine_dir trt_engines --precision fp32 --config config/tts.json

Citations

@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

Hebrew G2P uses renikud. Thanks to thewh1teagle.

License

MIT

Voice cloning and responsibility

This software can produce speech that mimics a reference voice. The maintainers and contributors are not responsible for what you do with it—compliance with law, consent from voice owners, and ethical use are entirely your responsibility. Do not use it to deceive, impersonate without permission, or infringe anyone’s rights.