Natiq-dz-ARG (ناطق)

Developed by Natiq Studio 🇩🇿

Natiq-dz-ARG is a production-grade speech synthesis and zero-shot voice cloning model tailored specifically for Algerian Darija (الدارجة الجزائرية) with natural French code-switching. It is built as a parameter-efficient fine-tuning (LoRA) adapter on top of bosonai/higgs-tts-3-4b.

The model synthesizes high-fidelity 24,000 Hz (24 kHz Mono) speech conditioned on target text and a short reference audio prompt.

---

Inference Guidelines & Reference Audio

The model uses zero-shot acoustic conditioning to replicate vocal timbre, cadence, and acoustic environment:

1. Reference Audio: Provide a clean, single-speaker audio recording between 2 and 20 seconds. The audio can be in WAV, MP3, or FLAC format (automatically converted to 24 kHz mono).
2. Reference Transcript: Providing an accurate verbatim transcript of what is spoken in the reference audio significantly improves synthesis fidelity and pronunciation.
3. Target Text: Enter the text you wish to synthesize in Algerian Darija (Arabic script, Latin transliteration, or natural French code-switching).
4. Precision: The model supports bfloat16 on modern GPUs (Ampere / Ada / Hopper) and float16 on NVIDIA T4 GPUs.

---

Python Usage

You can run the model on any CUDA GPU (≥16 GB VRAM) with transformers and peft. The model architecture code ships in this repo under model/:

import sys, torch, soundfile as sf
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer
from peft import PeftModel

# Download repository files and import custom model architecture
repo = snapshot_download("Natiq-Studio/Natiq-dz-ARG", ignore_patterns=["*.ipynb"])
sys.path.insert(0, repo)
from model.modeling import HiggsMultimodalQwen3ForConditionalGeneration

dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
tokenizer = AutoTokenizer.from_pretrained("bosonai/higgs-tts-3-4b")
base = HiggsMultimodalQwen3ForConditionalGeneration.from_pretrained("bosonai/higgs-tts-3-4b", torch_dtype=dtype)
base.prepare_inputs_for_generation = lambda *a, **k: {}  # PEFT compatibility shim
model = PeftModel.from_pretrained(base, repo).to("cuda").eval().base_model.model
model.audio_head.weight = model.audio_embedding.weight  # tied weights

ref, sr = sf.read(f"{repo}/sample_voices/ref_tech_developer.wav", dtype="float32")
with torch.inference_mode():
    wav = model.generate_speech(
        "واش راك خويا، واش راهم الدار؟",
        tokenizer,
        reference_audio=torch.from_numpy(ref).unsqueeze(0),
        reference_sample_rate=sr,
        reference_text="Développeur؟ لا لا لا، physicien de base، كيما كان يقول واحد le prof تاع physique، qui peut le plus peut le moins، بلا ما نعايروا حتى واحد.",
        temperature=0.7, top_p=0.95, max_new_tokens=768,
    )
sf.write("natiq_output.wav", wav.float().cpu().numpy().squeeze(), 24000)

Version

Natiq-dz-ARG v1.0
Released: October 2026

Training & Evaluation

An article covering the training procedure, speaker balancing, experiments: Technical article: https://natiq.studio/articles/natiq-v1-algerian-ai-speech-model-en/

Responsible Use

Natiq-dz-ARG supports zero-shot voice cloning from short reference audio.

Users are responsible for ensuring that they have the appropriate rights and consent to use any reference voice provided to the model.

The model must not be used for:

• non-consensual voice cloning
• impersonation or identity misrepresentation
• fraud, scams, or social engineering
• deceptive or malicious synthetic audio
• generating content intended to falsely attribute speech to another person

Natiq Studio does not endorse uses that violate applicable laws, individual rights, or consent requirements.

Commercial Licensing

The Natiq-dz-ARG adapter weights are released under CC BY-NC 4.0 for non-commercial use.

For commercial use of:

• the Natiq-dz-ARG adapter
• the Algeria-TTS-dataset
• custom Algerian Darija fine-tuning
• custom voice models
• hosted inference or enterprise API services

please contact Natiq Studio.

Commercial licensing from Natiq Studio covers only assets and services owned or provided by Natiq Studio.

Natiq-dz-ARG is derived from bosonai/higgs-tts-3-4b, and use of the underlying base model remains subject to the applicable Boson AI license. Users are responsible for ensuring compliance with both the Natiq Studio licensing terms and the upstream base-model license.

Training Corpus

• Training Dataset: Fine-tuned on the Algeria-TTS-dataset, a proprietary ~303-hour curated Algerian Darija speech collection. Access to the training dataset is gated and subject to separate commercial licensing agreements with Natiq Studio.

---

Contact & Commercial Licensing

For commercial licensing, custom voice cloning models, or enterprise speech API access:

• Website: Natiq Studio
• Email: abdelhak.kadouci@ensia.edu.dz

Citation

If you use Natiq-dz-ARG in research, experiments, or academic work, please cite:

@misc{kadouci2026natiq,
  author       = {Abdelhak Kadouci},
  title        = {Natiq-dz-ARG: Algerian Darija Text-to-Speech},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Natiq-Studio/Natiq-dz-ARG}}
}

---