Messam174/Sada-TTS-SA-v1

Sada TTS SA v1

A locally exported Arabic XTTS checkpoint fine-tuned for Arabic speech synthesis and inference in a project-specific voice pipeline.

This model package contains the exported inference artifacts for the trained Arabic voice: the XTTS configuration, weights, vocabulary, and speaker-conditioning latent files. It is intended for local testing, evaluation, demos, and in-house deployment workflows.

Model summary

• Language: Arabic
• Base model: Coqui XTTS v2
• Fine-tuning status: project-specific Arabic adaptation
• Output sample rate: 24 kHz
• Audio format: mono PCM16
• Inference target: local generation and streaming playback
• Speaker conditioning: included via exported speaker latent artifacts
• Runtime environment used for benchmarking: CUDA on NVIDIA GeForce RTX 3060

Intended use

This model is intended for:

• Arabic speech synthesis and voice generation
• local experimentation and prototyping
• evaluation and demo generation
• custom voice applications in controlled project environments

This model is not presented as a general-purpose public benchmark model or as a replacement for the upstream XTTS base model.

Evaluation results

The model was benchmarked locally on CUDA using a deterministic internal Arabic benchmark manifest (32 prompts, 3 repetitions per prompt). This is an internal curated benchmark and not a standardized public leaderboard benchmark.

At a glance

Intelligibility proxy. Normalized WER/CER macro: 0.352 / 0.137.	Latency profile. Warm mean latency: 1.599 s (median 1.534 s, p90 2.428 s).	Real-time behavior. Mean RTF 0.291 and mean RTFx 3.451 on RTX 3060.

Benchmark configuration

Field	Value
Run timestamp (UTC)	2026-10-25
Prompt set	Internal curated benchmark (32 prompts)
Repetitions per prompt	3
Warm-up runs	2
ASR model for proxy scoring	faster-whisper small
Device	NVIDIA GeForce RTX 3060 (12 GB), CUDA 12.8
Successful generations	96/96

Latency and throughput

Metric	Value
Model initialization time	11.373 s
First-inference latency	0.985 s
Warm inference latency (mean / median / p90 / std)	1.599 / 1.534 / 2.428 / 0.707 s
Total request latency (mean / median / p90 / std)	2.238 / 2.127 / 3.305 / 0.928 s
RTF (mean / median / p90 / std)	0.291 / 0.289 / 0.302 / 0.016
RTFx (mean / median / p90 / std)	3.451 / 3.457 / 3.611 / 0.164
Characters per second (mean / median / p90 / std)	42.858 / 43.299 / 53.569 / 8.613
Words per second (mean / median / p90 / std)	7.475 / 7.360 / 9.243 / 1.577

Intelligibility proxy metrics (ASR-based)

Metric	Macro average	Corpus aggregate
WER (normalized)	0.351839	0.360000
CER (normalized)	0.136596	0.135629

Availability of additional metrics

Metric	Status	Notes
Streaming TTFA / first-chunk latency	Not measured	Offline full-waveform synthesis path was benchmarked
Speaker similarity	Not available in this run	speechbrain not installed in benchmark environment
Predicted MOS	Not run	No validated Arabic MOS predictor configured
Human MOS	Not run	Template prepared; human study pending

Audio examples

These are synthesized Arabic samples generated by the exported model. They are not reference recordings and are included for demonstration purposes only.

Example 1

Text: "مرحبا، كيف الحال اليوم؟"

Example 2

Text: "هل يمكن أن تخبرني متى يبدأ الموعد وكيف يمكنني الاستعداد له؟"

Example 3

Text: "يعطيك العافية، هذا اليوم حلو جدًا ومرتاح في الرياض."

Example 4

Text: "سأعطيك ملخصًا مختصرًا عن الخطة الحالية: أولًا، نراجع جميع المتطلبات؛ ثانيًا، نحدد الأولويات؛ ثالثًا، نبدأ التنفيذ بطريقة منظمة مع متابعة دقيقة للنتائج في كل مرحلة."

Example 5

Text: "السلام عليكم، أنا فهد من قسم خدمة العملاء، وش أقدر أساعدك فيه اليوم؟"

Example 6

Text: "في واحد قالي على عاده عندكم بالزواجات و أنصدمت منها، مدري إذا صح ولا لا، بس ما يحضرون زواج الاخت"

Example 7

Text: "الحين ابي اطلع قرض شخصي ورحت اكثر من فرع يجيني رفض على القرض وشركتي معتمدة عندكم وسجلي في سمة 602 ولا عندي متعثرات والموظفين مايعرفون سبب الرفض، اذا الموظف مايعرف السبب مين يعرف السبب؟"

Example inference

from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts
import torch

model_dir = "./"
cfg = XttsConfig()
cfg.load_json(f"{model_dir}/config.json")
model = Xtts.init_from_config(cfg)
model.load_checkpoint(
    cfg,
    checkpoint_path=f"{model_dir}/model.pth",
    vocab_path=f"{model_dir}/vocab.json",
    use_deepspeed=False,
)
model.to("cuda")
model.eval()

lat = torch.load(f"{model_dir}/speaker_latents.pt", map_location="cpu")
gpt_cond = lat["gpt_cond_latent"].to("cuda")
spk = lat["speaker_embedding"].to("cuda")

wav = model.inference(
    "الحين ابي اطلع قرض شخصي ورحت اكثر من فرع يجيني رفض على القرض وشركتي معتمدة عندكم وسجلي في سمة 602 ولا عندي متعثرات والموظفين مايعرفون سبب الرفض، اذا الموظف مايعرف السبب مين يعرف السبب؟",
    "ar",
    gpt_cond,
    spk,
    temperature=0.55,
    repetition_penalty=4.5,
    speed=1.0,
    enable_text_splitting=False,
)["wav"]

Notes and limitations

• This package is an exported inference artifact, not a full training dataset release.
• Voice identity is tied to the speaker-conditioning files shipped with this checkpoint.
• Audio quality may vary by text length, punctuation, and prompt style.
• The model should be considered a project-specific voice adaptation and not a universal Arabic voice benchmark.
• For production or public release decisions, independent rights review is still recommended for any training data and downstream deployment context.

License

This model package inherits the base model licensing constraints of Coqui XTTS v2. The relevant license is the Coqui Public Model License (CPML), which applies to the base model and to distribution choices derived from it.

This repository does not redistribute a public training dataset or a YouTube archive. It contains only the exported model artifact and metadata for local inference and evaluation. Users must ensure that their own training data, deployment context, and downstream distribution comply with all applicable rights and licensing requirements.

Citation

If you use this model in research or demos, please cite the project and base model stack, and clearly disclose the model's fine-tuned, project-specific nature.