Configuration Parsing Warning:Invalid JSON for config file config.json

XTTS-v2 Egyptian Arabic (Fine-tuned)

A fine-tuned version of Coqui XTTS-v2 adapted to Egyptian Arabic speech. The model was trained on xtts-egyptian-dataset-v2, a single-speaker Egyptian Arabic corpus, and supports reference-audio-based voice conditioning like the original XTTS-v2.

⚠️ Research and experimental use only. Not for production. This is an experimental checkpoint. It has not been evaluated for robustness, safety, or reliability, and it must not be used in production systems, commercial products, or any setting where errors or misuse could cause harm.


Model Summary

Property Value
Base model coqui/XTTS-v2
Task Text-to-Speech (with reference-audio conditioning)
Target language Egyptian Arabic (use language code ar)
Fine-tuning data abdullah-khaled/xtts-egyptian-dataset-v2 (single speaker, 22.05 kHz)
Output sample rate 24,000 Hz
Framework Coqui TTS (XttsConfig / Xtts)
Status Experimental research checkpoint

🔊 Audio Sample

Listen to a sample generated by this model:

Input text: الماء مهم جداً لجسم الإنسان.

⬇️ Download output.wav

Generated with the settings listed in Recommended generation settings. The sample is provided for research demonstration only and quality may vary between runs and reference clips.


Repository Files

File Description
model.pth Fine-tuned model weights
config.json XTTS model configuration
vocab.json Tokenizer vocabulary
output.wav Audio sample generated by the model

Quick Start

1. Install dependencies

pip install coqui-tts torch torchaudio huggingface_hub

coqui-tts is the maintained fork of Coqui TTS. The original TTS package also exposes the same TTS.tts.models.xtts modules, but may need older dependency versions.

2. Run inference

You need a reference audio file (reference.wav): clean, mono, roughly 6–30 seconds of speech. The model uses it to compute the conditioning latents.

import os
import torch
import torchaudio
from huggingface_hub import snapshot_download
from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts

# Download the model (log in first with `huggingface-cli login`
# if the repository is private)
d = snapshot_download("abdullah-khaled/xtts-egyptian-arabic", token=True)

# Device
device = "cuda" if torch.cuda.is_available() else "cpu"
if device == "cpu":
    torch.set_num_threads(os.cpu_count())  # XTTS on CPU is slow

# Load model
config = XttsConfig()
config.load_json(f"{d}/config.json")

model = Xtts(config)
model.load_checkpoint(
    config,
    checkpoint_dir=d,
    checkpoint_path=f"{d}/model.pth",
    vocab_path=f"{d}/vocab.json",
    use_deepspeed=False,
)

# NOTE: .to() and .eval() return None for XTTS -- do not chain or reassign.
model.to(device)
model.eval()

# Conditioning latents from a reference clip
gpt_cond_latent, speaker_embedding = model.get_conditioning_latents(
    audio_path=["reference.wav"],
    gpt_cond_len=6,
    max_ref_length=30,
    sound_norm_refs=False,
)

# Synthesize
out = model.inference(
    text="الماء مهم جداً لجسم الإنسان.",
    language="ar",
    gpt_cond_latent=gpt_cond_latent,
    speaker_embedding=speaker_embedding,
    temperature=0.65,
    repetition_penalty=5.0,
    top_k=50,
    length_penalty=1.0,
    enable_text_splitting=True,  # auto-splits long text
)

# Save (XTTS-v2 outputs 24 kHz audio)
torchaudio.save("output.wav", torch.tensor(out["wav"]).unsqueeze(0), 24000)

Recommended generation settings

These are the settings used in the example above and a reasonable starting point:

Parameter Value Notes
language "ar" XTTS has no dedicated Egyptian code; ar is used
temperature 0.65 Lower = more stable, higher = more expressive
repetition_penalty 5.0 Helps prevent looping/repeated syllables
top_k 50
length_penalty 1.0
enable_text_splitting True Recommended for long inputs

Training Details

  • Method: Fine-tuning of the pretrained XTTS-v2 checkpoint on Egyptian Arabic clips.
  • Data: Sentence-level clips of up to 12 seconds, 22,050 Hz mono, with LLM-generated (not fully human-verified) transcripts. Full details are in the dataset card.
  • Hyperparameters, epochs, hardware: Add your training details here (e.g., learning rate, batch size, number of epochs/steps, GPU).

Intended Use & Restrictions

✅ Permitted

  • Academic and non-commercial research on Arabic / Egyptian-dialect speech synthesis
  • Personal experimentation, demos, and benchmarking
  • Educational use and reproducibility studies

❌ Not permitted

  • Production deployment of any kind
  • Commercial use, including paid products, APIs, or monetized content
  • Cloning or imitating a real person's voice without their explicit consent
  • Impersonation, fraud, deepfakes, misinformation, or harassment
  • Presenting synthetic speech as genuine human speech without disclosure

Because this model is built on XTTS-v2, it also inherits the terms of the Coqui Public Model License (CPML), which restricts commercial use. Please review it before using or redistributing this model.


Limitations & Known Issues

  • Experimental quality: Output may contain mispronunciations, unnatural prosody, skipped or repeated words, hallucinated sounds, or artifacts, especially on long, rare, or out-of-domain text.
  • Single-speaker fine-tuning data: The training data has one speaker, so the model may be biased toward that voice and may not clone other voices as faithfully as the original XTTS-v2.
  • Imperfect transcripts: Training transcripts were machine-generated and may contain errors, which can transfer to the model's pronunciation or text handling.
  • Dialect coverage: Egyptian Arabic has no standard orthography. Input text style (colloquial spelling vs. Modern Standard Arabic), diacritics, numbers, and Arabic/English code-switching can strongly affect quality.
  • No formal evaluation: No objective (e.g., CER/WER via ASR round-trip, speaker similarity) or subjective (MOS) evaluation is reported for this checkpoint.
  • Reference audio sensitivity: Output quality depends heavily on the cleanliness and length of reference.wav.
  • Inference speed: Running on CPU is slow; a CUDA GPU is strongly recommended.

Ethical Considerations

Text-to-speech and voice-cloning models can be misused to deceive people or imitate individuals. By using this model you agree to:

  • Obtain consent from any person whose voice you use as a reference
  • Clearly label synthetic audio as AI-generated where appropriate
  • Comply with all applicable laws and platform policies on synthetic media

Acknowledgements


Citation

@misc{khaled2026xttsegyptianmodel,
  author       = {Abdullah Khaled},
  title        = {XTTS-v2 Egyptian Arabic (Fine-tuned)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/abdullah-khaled/xtts-egyptian-arabic}}
}

Contact

Maintained by Abdullah Khaled. Questions and issues are welcome via the Community tab.

Downloads last month
62
Inference Examples
Examples
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for abdullah-khaled/xtts-egyptian-arabic

Base model

coqui/XTTS-v2
Finetuned
(84)
this model