Darija OmniVoice — Kore

A full fine-tune of k2-fsa/OmniVoice for Moroccan Arabic (Darija), targeting the Kore voice.

This checkpoint specializes OmniVoice toward clearer Darija pronunciation, more natural rhythm and prosody, and a consistent speaker — rather than relying on the base model's generic multilingual Darija coverage alone.

For the full write-up (data, training setup, design choices, and evaluation notes), see:

Training a voice

Quick facts


Base model	`k2-fsa/OmniVoice`
Language ID	`ary`
Target voice	Kore
Steps	5,000
Examples	8,400 train / 187 val
Precision	BF16
Hardware	1× NVIDIA H100
Type	Full fine-tune
Emotion / instruction conditioning	Disabled
Audio tokenizer	`eustlb/higgs-audio-v2-tokenizer`

Experimental research release. No MOS / WER / speaker-similarity numbers yet — listen before making strong quality claims.

Usage

pip install omnivoice
# or: git clone https://github.com/k2-fsa/OmniVoice.git && cd OmniVoice && pip install -e .

import torch
import soundfile as sf
from omnivoice import OmniVoice

model = OmniVoice.from_pretrained(
    "ai-ssam/darija-omnivoice-kore-v1",
    device_map="cuda:0",
    dtype=torch.float16,
)

audio = model.generate(
    text="السلام عليكم، كيدايرين؟ نتمنى تكونو بخير.",
    language="ary",
)

sf.write("darija.wav", audio[0], model.sampling_rate)

omnivoice-infer \
  --model ai-ssam/darija-omnivoice-kore-v1 \
  --text "السلام عليكم، كيدايرين؟" \
  --language ary \
  --output darija.wav

Passing language="ary" usually gives a clearer conditioning signal. The audio tokenizer is fetched automatically if not present locally.

License

CC BY-NC 4.0 — commercial use is not permitted.

Upstream OmniVoice code is Apache-2.0; the pretrained weights are CC-BY-NC, so this derivative stays non-commercial. See k2-fsa/OmniVoice.

Do not use for deceptive impersonation, fraud, or unauthorized voice cloning.

Citation

@article{zhu2026omnivoice,
  title={OmniVoice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models},
  author={Zhu, Han and Ye, Lingxuan and Kang, Wei and Yao, Zengwei and Guo, Liyong and Kuang, Fangjun and Han, Zhifeng and Zhuang, Weiji and Lin, Long and Povey, Daniel},
  journal={arXiv preprint arXiv:2604.00688},
  year={2026}
}