🎙️ Arabic TTS Arena

Blind A/B voting platform for Arabic text-to-speech models. Users hear two anonymous outputs, pick the better one, and an ELO system ranks all models on a public leaderboard.

We welcome all Arabic TTS models — open-source or proprietary.

---

Add Your Model

Open-source model

1. Create models/your_model.py with its own image:

import modal
from models import BaseTTSModel, arena_cls, register_model
from app import LOCAL_MODULES

# Each model defines its own image — install only what you need
your_image = (
    modal.Image.from_registry("nvidia/cuda:12.8.0-devel-ubuntu24.04", add_python="3.12")
    .apt_install("ffmpeg", "libsndfile1")
    .uv_pip_install("torch>=2.0.0", "numpy", "soundfile", "your-model-package")
    .add_local_python_source(*LOCAL_MODULES)
)

@register_model
@arena_cls(image=your_image, gpu="T4", scaledown_window=300,
           secrets=[modal.Secret.from_name("hf-ar-tts-arena")])
class YourModel(BaseTTSModel):
    model_id = "your_model"
    display_name = "Your Model"
    model_url = "https://huggingface.co/your-org/your-model"

    @modal.enter()
    def load_model(self):
        ...  # load weights once on container start

    @modal.method()
    def synthesize(self, text: str) -> dict:
        try:
            wav = ...  # numpy array
            return self.success_response(self.audio_to_base64(wav, 24000), 24000)
        except Exception as e:
            return self.error_response(e)

2. Open a PR.

API-based / closed-source model

1. Copy models/example_api_model.py and adapt it. Lightweight image, no GPU.
2. Open a PR with your model code.
3. DM the maintainer your API keys. We store them in an encrypted vault (never in git).

Rules

• Must support Arabic text input
• Return audio via success_response() / errors via error_response()
• Unique model_id (lowercase, underscores)
• Inference under 30 seconds

CI auto-deploys to Modal on merge.

License

Apache 2.0