livekit-turn-detector-arabic

LiveKit Turn Detector - Arabic

Fine-tuned Arabic End-of-Utterance (EOU) detection model for LiveKit voice agents.

Model Details

• Base Model: livekit/turn-detector (Qwen2-0.5B)
• Fine-tuning Method: LoRA (rank=32, alpha=64)
• Dataset: 57,475 Arabic EOU samples
• Languages: Arabic (ar, ar-SA, ar-EG, Gulf dialects)
• Training: 3 epochs on T4 GPU (~20-30 minutes)
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Use Cases

• Real-time Arabic voice agents with LiveKit
• Turn-taking detection in Arabic conversations
• End-of-utterance detection for Gulf Arabic dialects
• Multilingual voice assistants with Arabic support

Usage

Basic Usage with Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model
model = AutoModelForCausalLM.from_pretrained("Moustafa3092/livekit-turn-detector-arabic")
tokenizer = AutoTokenizer.from_pretrained("Moustafa3092/livekit-turn-detector-arabic", trust_remote_code=True)

# Predict EOU
def predict_eou(text: str) -> float:
    formatted = f"<|im_start|>user\n{text}"
    inputs = tokenizer(formatted, return_tensors="pt", add_special_tokens=False)

    with torch.no_grad():
        logits = model(**inputs).logits[0, -1, :]
        probs = torch.softmax(logits, dim=-1)
        eou_prob = probs[tokenizer.convert_tokens_to_ids("<|im_end|>")]

    return eou_prob.item()

# Test
print(predict_eou("شكرا جزيلا"))  # Should be high (complete)
print(predict_eou("اممممم"))      # Should be low (incomplete)

With LiveKit Voice Agents

1. Export to ONNX:

   # Convert model to ONNX format for production
   # See deployment documentation
   

2. Use in LiveKit Agent:

   from livekit.agents import WorkerOptions, cli
   from livekit.plugins import turn_detector
   
   # Configure with your fine-tuned model
   turn_detector.configure(model_path="path/to/model.onnx")
   

Training Data

Dataset Composition (57,475 samples)

Complete Utterances (EOU): 20,194 (35.1%)

• CSV data: 19,432 samples
• Edge case closures: 762 samples
• Examples: "شكرا" (thank you), "تمام" (perfect), "مع السلامة" (goodbye)

Incomplete Utterances (non-EOU): 37,281 (64.9%)

• Generated variants: 36,610 samples
• Edge case hesitations: 671 samples
• Examples: "اممممم" (ummm...), "يعني" (you know...), "بس" (but...)

Edge Cases (1,433 samples)

Hesitations (Non-EOU) - Speaker thinking, will continue:

• اممممم, يعني, خلاص بس, طيب و

Closures (EOU) - Complete short responses:

• شكرا, تمام, نعم, لا, مع السلامة

Performance

• Accuracy: >90% on Arabic edge cases
• Training Time: ~20-30 minutes (T4 GPU)
• Memory: ~6 GB VRAM during training
• Inference: Real-time compatible
• Model Size: 494 MB (full model)

Model Architecture

• Architecture: Qwen2-0.5B (Causal LM)
• Parameters: ~500M total
• Fine-tuned: LoRA adapters merged
• Context Length: Supports conversation context
• Output: EOU probability via <|im_end|> token

Limitations

• Optimized for Modern Standard Arabic and Gulf dialects
• May need additional fine-tuning for other Arabic dialects
• Requires sufficient context for accurate predictions
• Best performance on conversational Arabic

Citation

If you use this model, please cite:

@misc{livekit-turn-detector-arabic,
  author = {Moustafa3092},
  title = {LiveKit Turn Detector - Arabic},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/Moustafa3092/livekit-turn-detector-arabic}
}

License

Apache 2.0

Acknowledgements

• Based on LiveKit Turn Detector
• Built with Transformers and PEFT
• Fine-tuned for Arabic language support

---

Developed by: Moustafa3092 Model Type: Turn Detection / End-of-Utterance Language: Arabic (ar) Base Model: livekit/turn-detector (v0.4.1-intl)