phoonnx_ar_miro_espeak
This is a text-to-speech voice model for Arabic. The voice is Miro, a male voice. It uses the phoonnx VITS architecture with espeak-ng for phonemization, exported to ONNX for CPU inference.
A newer version of this voice is available: Phoonnx/phoonnx_ar_miro_espeak_V2.
Model details
| Field | Value |
|---|---|
| Repository | Phoonnx/phoonnx_ar_miro_espeak |
| Voice id | Not in the phoonnx voice index |
| Languages | ar |
| Architecture | VITS |
| Text input | espeak phonemes, IPA alphabet |
| Punctuation | Trained without punctuation. phoonnx removes it before synthesis. Each clause break reached training as the line-break token that espeak-ng prints. phoonnx sends no such token at synthesis. |
| Sample rate | 22050 Hz |
| Speakers | 1 |
| Training code | phoonnx, version stamp 0.0.0 (the version string of the training code; a development build carries the stamp of an earlier release) |
| Phonemizer | espeak-ng, version not recorded. If a released espeak-ng was used, it was at most 1.52.0 (released 2024-12-12), the newest release before the model was first uploaded on 2025-09-21. |
| Training data | Phoonnx/tts-train-synthetic-miro_ar-SA |
| Base model | Not recorded |
| Training checkpoint | epoch=1127-step=148896.ckpt, in this repository for fine-tuning |
| Trained by | TigreGotico Lda |
| Licence | CC BY-NC-ND 4.0 |
How to use
The phoonnx voice index does not list this voice. Install phoonnx:
pip install --pre phoonnx
The phonemes come from the espeak-ng program, so install it too (for example
apt install espeak-ng). Without it, pip install --pre "phoonnx[espeak]" gives
phoonnx the espyak port of espeak-ng. Some espyak phonemes differ from those of
espeak-ng, so the voice can say some words wrong with espyak.
Then load the voice from its files:
import wave
from huggingface_hub import hf_hub_download
from phoonnx.voice import TTSVoice
model = hf_hub_download("Phoonnx/phoonnx_ar_miro_espeak", "miro_ar.onnx")
config = hf_hub_download("Phoonnx/phoonnx_ar_miro_espeak", "miro_ar.json")
voice = TTSVoice.load(model, config)
with wave.open("output.wav", "wb") as wav_file:
voice.synthesize_wav("ุตุจุงุญ ุงูุฎูุฑุ ููู ุญุงููุ", wav_file)
Training data
Trained on Phoonnx/tts-train-synthetic-miro_ar-SA.
Related links
- Training-data collection: https://huggingface.co/collections/TigreGotico/synthetic-tts-datasets
- phoonnx: https://github.com/TigreGotico/phoonnx
- voiceclonnx: https://github.com/TigreGotico/voiceclonnx
- TigreGotico: https://tigregotico.pt
Ownership and licensing
Miro and Dii are the recorded voices of two real people. The voices, the training data, and this model belong to TigreGotico Lda (https://tigregotico.pt).
This model is licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 (CC BY-NC-ND 4.0). You may use and share this model for non-commercial purposes only. Give attribution to TigreGotico Lda. Do not modify, adapt, or build new models from this model. Do not use the voice identity to create derivative voices.
For commercial use, for derivative models, or for any other license of the Miro or Dii voice identity, contact TigreGotico Lda.