Dezheen89/omnivoice-badini

OmniVoice Badini (Central/Northern Kurdish fine-tune)

This is a fine-tune of k2-fsa/OmniVoice -- a massively multilingual zero-shot text-to-speech model -- on Badini Kurdish (kmr) speech data, trained for 10,000 steps. It performs zero-shot voice cloning: give it a short reference audio clip and the text you want spoken, and it generates that text in the reference speaker's voice.

Files

• model.safetensors -- fine-tuned weights
• config.json -- model config (OmniVoice + Qwen3 LLM backbone)
• tokenizer.json, tokenizer_config.json -- tokenizer
• chat_template.jinja -- chat template used for prompting
• training_config.json -- training hyperparameters, provided for reference only (not required for inference)
• prompt-hayfa.wav / prompt-hayfa.txt -- an original (real, not model-generated) reference voice clip and its transcript, ready to use as ref_audio / ref_text below

Note: optimizer/scheduler/RNG state from training were intentionally excluded -- they're only needed to resume training, not for inference.

Usage

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128
pip install omnivoice soundfile webdataset accelerate huggingface_hub s3prl

import torch, soundfile as sf
from omnivoice import OmniVoice
from huggingface_hub import hf_hub_download

REPO_ID = "Dezheen89/omnivoice-badini"

model = OmniVoice.from_pretrained(REPO_ID, device_map="cuda:0", dtype=torch.float16)

# prompt-hayfa.wav/.txt are bundled in this repo as a ready-to-use example
# reference voice (a real recording, with its real transcript) -- fetch them
# alongside the model:
ref_audio = hf_hub_download(REPO_ID, "prompt-hayfa.wav")
ref_text = open(hf_hub_download(REPO_ID, "prompt-hayfa.txt"), encoding="utf-8").read().strip()

# Zero-shot voice cloning: give a short reference clip (+ its transcript) and
# the text you want spoken in that voice.
audio = model.generate(
    text="سەروکاتیا زانکۆیا دهۆک رادگەهینیت کو دێ رابیت ب راگەهاندنا پوستێ رێڤەبەرێ سنتەرێ ڤەکولینان ل کولیژا کارگێری و ئابووری، هەر ماموستایەکێ حەز هەبیت مافێ پێشکێشکرنێ یێ هەی بو پوستێن دیارکری برێیا (لینکێ) پوستێن هەڤپێچ کری.",
    ref_audio=ref_audio,
    ref_text=ref_text,
    language="kmr",
)
sf.write("output.wav", audio[0], 24000)  # output is 24kHz

Verified end-to-end on 2026-09-09: from_pretrained loading the Hub repo ID directly, downloading prompt-hayfa.wav/.txt fresh via hf_hub_download, and generate() all run successfully with this exact code.

License

Base model (k2-fsa/OmniVoice) is released under Apache-2.0. Note the base model's own model card flags that its bundled Higgs Audio tokenizer is under a separate, non-Apache license -- check https://huggingface.co/k2-fsa/OmniVoice for the tokenizer's LICENSE file before redistributing.

Training data

Fine-tuned on Badini Kurdish (kmr) speech data. [Add dataset details/size/hours here.]