Generating natural speech in Algerian Darija is harder than standard benchmarks suggest. Off-the-shelf multilingual models run into two problems immediately: they either force Algerian phrasing into Classical Arabic phonetics, sounding like a formal news broadcast, or they fail on everyday French loanwords, turning phrases like "la livraison" or "facture" into distorted audio.
Natiq-dz-ARG is an open-weight text-to-speech model built to address this problem directly. We fine-tuned bosonai/higgs-tts-3-4b, a 3.4-billion-parameter speech foundation model, using Low-Rank Adaptation (LoRA). The resulting model outputs 24 kHz audio, handles conversational code-switching, and supports zero-shot voice cloning in Algerian Darija.
We released the model openly on Hugging Face so researchers and builders can inspect the weights and run their own experiments. Here is the engineering story behind the project, including the initial roadblocks that forced us to rethink our approach.
Why single-speaker training overfitted
Our initial setup was straightforward. We collected clean studio recordings from a single speaker, segmented the audio into thousands of short clips, and trained a small acoustic model on that single voice.
The experiment ran into immediate trouble:
The model memorized that speaker's pitch, speed, and vowel lengths almost word for word. On sentences outside the training set, it stumbled, dragged out syllables, or hallucinated phantom vowels.
It also broke down on French loanwords. A single speaker cannot naturally cover the full range of everyday vocabulary. When the model encountered common French terms, it forced them through Arabic consonant rules, creating jarring artifacts.
Finally, voice cloning failed completely. Because the model had only ever heard one voice, it conflated dialect pronunciation with speaker identity. Conditioning it on a new target voice simply reproduced the cadence and tone of the original speaker.
Training on one voice gave us an inflexible audio reader rather than a model that genuinely understood the dialect.
Shifting to multi-speaker data and 300+ hours of audio
To fix the overfitting, we changed data strategies completely. We expanded our training corpus to more than 300 hours of conversational Algerian audio across over 100 different speakers.
That shift resolved the earlier issues:
We included speakers across multiple regions (Algiers, Oran, Constantine, and southern provinces) with varied ages and styles. The audio covered technical discussions, commerce, casual dialogue, and narrative storytelling.
We also used Gemini Flash to clean the transcripts. In raw speech recognition output, French words were frequently transcribed as broken Arabic letters (for instance, writing "physicien" as "فيزيسيان"). Gemini Flash corrected those terms back to their original Latin spelling within the Arabic sentence. This taught the model when to preserve French phonetics without breaking the rhythm of the surrounding Darija words.
Why a 3.4B foundation model and LoRA
Early in the project, we tested smaller architectures under 500 million parameters. They struggled with memory capacity. Algerian Darija requires a model to maintain two vocabularies simultaneously: Arabic phonology with dialectal consonants like the voiced velar stop /g/ (الڨاف), and French commercial terms.
Moving to bosonai/higgs-tts-3-4b provided three main advantages:
Contextual prediction without manual diacritics. The 3.4B parameter transformer has enough capacity to infer whether a consonant is vocalized or silent based on sentence context. Users do not need to add vowel marks (Harakaat) to their text.
Stable adaptation with LoRA. Rather than retraining all weights, which risks catastrophic forgetting, we adapted low-rank matrices targeting dialectal phonetics. This retained the underlying model's stability while teaching it local cadence.
24 kHz output. The neural vocoder outputs at 24,000 Hz mono WAV, which removes the raspy, metallic buzz typical of smaller lightweight web synthesizers.
How zero-shot voice cloning works in Natiq-dz-ARG
Instead of fine-tuning a separate model for every character, Natiq-dz-ARG uses reference conditioning pairs:
- You supply a short audio sample of any speaker (between 3 and 10 seconds).
- The network extracts a speaker embedding representing vocal tract resonance and pitch.
- The model applies that embedding to the input text while relying on its LoRA weights for dialectal pronunciation.
- The output is a new voice speaking natural Algerian Darija, maintaining the vocal identity of the provided sample.
Testing on real-world scripts
We tested the model on everyday production tasks in Algeria:
For e-commerce video ads (COD), the model handles commercial phrases like "التوصيل لـ 58 ولاية" and "الدفع عند الاستلام" smoothly, letting store owners test multiple video hooks without booking studio sessions.
For short-form content, it produces fast-paced narration suitable for TikTok and Instagram Reels.
On dedicated GPUs, generation latency sits under 500 milliseconds for standard paragraphs, making the model fast enough for interactive voice pipelines.
How to test and run Natiq-dz-ARG
Natiq-dz-ARG is an open-source project and is not deployed as a proprietary engine on the web platform. It is hosted directly on Hugging Face, where anyone can experiment with it for free:
Try the interactive demo on Hugging Face Spaces: abdelhak-k/Natiq-dz-ARG Space. You can paste your own Darija text, upload a short reference voice, and generate audio directly in your browser.
Inspect the model weights on Hugging Face: abdelhak-k/Natiq-dz-ARG. Developers and researchers can download the weights and integrate them into local PyTorch pipelines.
Frequently Asked Questions
Is Natiq-dz-ARG an open-source model?
Yes. The model weights are openly accessible on Hugging Face for non-commercial research, testing, and community development.
Does the model require vowel diacritics (Harakat)?
No. Natiq-dz-ARG infers correct vowel lengths and silent consonants directly from sentence context. Standard commas and periods are sufficient to control breathing pauses.
Can the model read numbers and prices?
For best results in commercial ads, spell out prices phonetically the way they are spoken in Darija (such as "مية وخمسين ألف" instead of raw digits).
Does Natiq-dz-ARG also support Modern Standard Arabic?
Yes. The model can synthesize Modern Standard Arabic with clean intonation, though its primary focus is colloquial Algerian Darija and bilingual code-switching.