---
license: mit
tags:
- text-to-speech
- matcha-tts
- ipa
- onnx
- hebrew
- english
---

# Matcha-TTS ONNX Models

This repo contains full end-to-end Matcha-TTS ONNX graphs with the HiFiGAN vocoder embedded.

Code:

https://github.com/thewh1teagle/Matcha-TTS/tree/feature/train-on-phonemes

Python wrapper package:

https://pypi.org/project/matcha-onnx/

## Files

- `matcha-english.onnx` - original pretrained English/LJSpeech Matcha-TTS exported to ONNX.
- `matcha-he-en.onnx` - Hebrew + English IPA fine-tuned single-speaker model exported to ONNX.
- `matcha-he-en.ckpt` - latest PyTorch Lightning checkpoint for the Hebrew + English model.
- `sample-he-en.wav` / `sample-he-en.m4a` - sample generated from `matcha-he-en.ckpt`.
- `sample-he-en.txt` - exact sample phonemes and synthesis settings.

Both ONNX files include metadata for:

- `matcha.symbols`
- `matcha.sample_rate`
- `matcha.n_timesteps`
- `matcha.embedded_vocoder`
- `matcha.n_spks`

## Usage

Use the `matcha-onnx` package. It reads the symbol table, sample rate, and step count from ONNX metadata.

```bash
pip install matcha-onnx
wget -O matcha.onnx https://huggingface.co/thewh1teagle/matcha-tts/resolve/main/matcha-english.onnx
```

```python
import soundfile as sf
from matcha_onnx import MatchaOnnx

tts = MatchaOnnx("matcha.onnx")
samples, sr = tts.create("Hello from Matcha TTS.", speed=0.95)
sf.write("audio.wav", samples, sr)
```

For the Hebrew + English model, input must be IPA phonemes, not raw Hebrew or English graphemes.

```bash
wget -O matcha-he-en.onnx https://huggingface.co/thewh1teagle/matcha-tts/resolve/main/matcha-he-en.onnx
```

```python
from matcha_onnx import MatchaOnnx

tts = MatchaOnnx("matcha-he-en.onnx")
samples, sr = tts.create("sˈimu lˈev nosʔˈim jekaʁˈim.", is_phonemes=True)
```

Audio config: mono, 22050 Hz.