---
language:
  - multilingual
tags:
  - text-to-speech
  - tts
  - hebrew
  - audio
  - fast-inference
  - onnx
license: mit
datasets:
  - notmax123/RanLevi40h
  - notmax123/SententicDataTTS
pipeline_tag: text-to-speech
---

# Blue ONNX — Text-to-speech inference

This repository is the **ONNX model bundle** for **[BlueTTS](https://github.com/maxmelichov/BlueTTS)**: fast Hebrew-first multilingual speech synthesis with [ONNX Runtime](https://onnxruntime.ai/) and optional NVIDIA TensorRT engines (see the GitHub repo).

**Project home (install, usage, examples, TensorRT):** [https://github.com/maxmelichov/BlueTTS](https://github.com/maxmelichov/BlueTTS)

**Try in the browser:** [Hugging Face Space — notmax123/Blue](https://huggingface.co/spaces/notmax123/Blue) · **Website:** [lightbluetts.com](https://lightbluetts.com/)

## Model description

BlueTTS targets **native Hebrew** quality (including *nikud* and disambiguation of common homographs) while staying fast enough for **real-time and batch** use. The reference codebase also supports **English, Spanish, Italian, German**, and **mixed-language** segments in one utterance.

Hebrew G2P at inference uses the **renikud** ONNX model (`model.onnx`); download it separately (see below).

## Related weights

| Repo | Purpose |
|------|--------|
| This repo (`notmax123/blue-onnx`) | **ONNX** checkpoints for `BlueTTS` inference |
| [`notmax123/blue`](https://huggingface.co/notmax123/blue) | **PyTorch / Safetensors** weights and stats for **training, finetuning, and exporting new voice JSON** |

## Download

Repo id is **case-sensitive**: `notmax123/blue-onnx`.

```bash
hf download notmax123/blue-onnx --repo-type model --local-dir ./onnx_models
wget -O model.onnx https://huggingface.co/thewh1teagle/renikud/resolve/main/model.onnx
```

Classic CLI equivalent:

```bash
huggingface-cli download notmax123/blue-onnx --repo-type model --local-dir ./onnx_models
```

## How to get started

Clone **[BlueTTS](https://github.com/maxmelichov/BlueTTS)**, run `uv sync`, place the ONNX bundle and `model.onnx` as in the [project README](https://github.com/maxmelichov/BlueTTS/blob/main/README.md), then use `src.blue_onnx.BlueTTS` or the `examples/` scripts.

Minimal pattern:

```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",
)
samples, sr = tts.synthesize("Hello, this is a short English sample from BlueTTS.", lang="en")
sf.write("output.wav", samples, sr)
```

*(Adjust `PYTHONPATH` / imports if you install the package instead of running from a clone.)*

## Speed (indicative)

Reported reference throughput on tuned setups (see project materials and [lightbluetts.com](https://lightbluetts.com/)); your hardware, batching, and TensorRT vs ONNX Runtime settings will change these numbers.

| Hardware | Approx. speed | ~1 h audio |
|----------|----------------|------------|
| NVIDIA RTX 3090 (GPU) | very high RTF | on the order of seconds |
| Typical CPU | strong RTF | on the order of minutes |
| Apple M1 class | strong RTF | on the order of a few minutes |

## Uses

- Hebrew and multilingual TTS from text
- Real-time or offline apps on CPU or GPU
- Audiobooks, accessibility, assistants, and broadcasting pipelines

## Citations

BibTeX and paper links are maintained in the **[BlueTTS README](https://github.com/maxmelichov/BlueTTS/blob/main/README.md)**.

## License

MIT — see the [BlueTTS repository](https://github.com/maxmelichov/BlueTTS).