🇲🇦 HADRA (هضرة) — Habibi-TTS Moroccan Darija

🌟 Overview

Jip7e/habibi-tts-doda-darija (also known as HADRA / هضرة) is an open-source speech synthesis model fine-tuned specifically for Moroccan Darija (الدارجة المغربية).

Unlike autoregressive language models that suffer from token looping, hallucinated pauses, and phonetic collapse on dialectal Arabic, this model uses a non-autoregressive Flow-Matching Diffusion Transformer (DiT) backbone. It generates 24 kHz studio-quality Moroccan speech with authentic cadence, natural prosody, and razor-sharp zero-shot voice cloning.

• Live Demo & Voice Playground: https://hadra-seven.vercel.app/
• Base Architecture: F5-TTS DiT Backbone (22 layers, 1024-dim, 16 heads, 336M parameters)
• Vocoder: Vocos 24 kHz Mel-Vocoder (charactr/vocos-mel-24khz)
• Training Dataset: 12,105 studio clips from Jip7e/doda-darija-cosyvoice2
• Speakers: 7 Balanced Native Moroccan Personas (4 Female, 3 Male)
• Available Formats:
  • model_ema.safetensors (1.3 GB - Recommended): Lightweight, pure EMA weights for low-latency production inference and serverless deployment.
  • model_last.pt (5.4 GB): Full checkpoint containing model, EMA, and optimizer states.

🏛️ The Three-Tier Architecture

To balance studio-grade acoustic fidelity with sub-second real-time conversational agents, HADRA implements a coordinated three-tier routing system:

1. ⚡ Tier 0: 0ms Pre-Synthesized Opus Store: 24 high-frequency Moroccan greetings, affirmations, and filler phrases pre-computed at NFE=32 with fuzzy matching (RapidFuzz ratio ≥ 90%) for 0.01ms response time.
2. 🎙️ Tier 1: F5-TTS DiT Quality Engine: Flow-matching DiT with Sway Sampling (sway_sampling_coef = -1.0) and Truncated CFG for custom sentences, studio voiceovers, and longform narrative reading.
3. 🚀 Tier 2: Realtime Piper VITS ONNX (In Progress): Single-pass feedforward ONNX model (RTF < 0.05, 30ms latency) dedicated to live Voice-to-Voice conversational turns on CPU.

🎙️ Golden Reference Prompts for the 7 Speakers

F5-TTS is conditioned by inpainting directly on the mel-spectrogram canvas. Using a clean 3–6s reference audio with a 100% verbatim transcript ensures stable pitch, crisp consonants (ح, خ, ق, ع), and zero phoneme skipping.

Here are the curated golden prompts from the DODa dataset:

🚀 Quickstart & Inference

1. Installation

git clone https://github.com/SWivid/F5-TTS.git
cd F5-TTS
pip install -e .
pip install soundfile huggingface_hub vocos accelerate safetensors

2. Python Inference (With Sway Sampling & AMX Acceleration)

import soundfile as sf
import torch
from huggingface_hub import hf_hub_download
from f5_tts.infer.utils_infer import load_model, load_vocoder, preprocess_ref_audio_text, infer_process
from f5_tts.model import DiT

# 1. Download Model & Vocab
REPO_ID = "Jip7e/habibi-tts-doda-darija"
ckpt_path = hf_hub_download(repo_id=REPO_ID, filename="model_ema.safetensors")
vocab_path = hf_hub_download(repo_id=REPO_ID, filename="vocab.txt")

# 2. Load Model & Vocos Vocoder
cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
model = load_model(DiT, cfg, ckpt_path, vocab_file=vocab_path)
vocoder = load_vocoder("vocos")

# 3. Reference Speaker Audio (e.g. Salma F1)
ref_audio = "path/to/F1.wav"
ref_text = "هوما مخبيين شي حاجة انا متيقن"

# 4. Target Text in Moroccan Darija
target_text = "السلام عليكم ورحمة الله! تبارك الله عليك، هاد النموذج دالذكاء الاصطناعي المغربي ناضي بزاف."

# 5. Synthesize with Sway Sampling (NFE=16 for speed + studio quality)
ra, rt = preprocess_ref_audio_text(ref_audio, ref_text)

with torch.autocast(device_type="cpu" if not torch.cuda.is_available() else "cuda", dtype=torch.bfloat16):
    wav, sr, _ = infer_process(
        ra, 
        rt, 
        target_text, 
        model, 
        vocoder, 
        speed=0.95, 
        nfe_step=16, 
        cfg_strength=2.0,
        sway_sampling_coef=-1.0
    )

sf.write("output_darija.wav", wav, sr)
print("✅ Saved output_darija.wav!")

⚙️ Recommended Inference Hyperparameters

📊 Training Details & Benchmarks

• Hardware: 2x NVIDIA Tesla T4 GPUs (PyTorch DDP via Accelerate).
• Epochs: 12 Full Epochs (984 optimizer steps, Batch size: 3200 frames/GPU, Grad Acc: 6).
• Optimizer: AdamW (lr = 1e-5, cosine decay with 500 warmup steps).
• Final Loss: 0.59 (smooth monotonic convergence from 2.84).
• CPU Inference Benchmarks (Intel Xeon Platinum Sapphire Rapids AMX-BF16):
  • NFE 8: RTF 2.65 (~11s)
  • NFE 16 + Sway (-1.0): RTF 3.40 (~14s) — Recommended Sweet Spot
  • NFE 32: RTF 10.60 (~44s)
  • 0ms Cache: RTF 0.00 (0.01ms)

📜 Citation

@misc{hadra_darija_tts_2026,
  title={HADRA: Open-Source Moroccan Darija Flow-Matching Voice Model},
  author={Hadra Open Initiative and Contributing Medical Student},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/Jip7e/habibi-tts-doda-darija}},
  note={Fine-tuned on DODa Moroccan Arabic Dataset (12,105 Studio Utterances)}
}

🙏 Acknowledgements

• AtlasIA DODa Team: For creating and open-sourcing the foundational Moroccan Darija speech corpus.
• SWivid / F5-TTS Team: For the non-autoregressive Flow Matching DiT architecture.
• Habibi-TTS Authors: For the dialectal Arabic foundation models.