livekit-turn-detector-arabic

Without EOU (VAD only):             With EOU:

User: "رقم جوالي صفر خمسة..."       User: "رقم جوالي صفر خمسة..."
       500ms silence                        500ms silence
VAD:   "Silence detected!"           VAD:   "Silence detected!"
AI:    "Got it! Your number           EOU:   P(end) = 0.08 -> LOW
        is 055..."                          "Incomplete phone number"
       INTERRUPTED!                  AI:    keeps listening...
                                     User: "...واحد ثلاثة اثنين خمسة"
                                     EOU:   P(end) = 0.92 -> HIGH
                                     AI:    "Got it! 0551325..."
                                            CORRECT!

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained("uselevers/livekit-turn-detector-arabic")
tokenizer = AutoTokenizer.from_pretrained("uselevers/livekit-turn-detector-arabic")

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()

print(predict_eou("شكرا جزيلا"))    # HIGH - complete
print(predict_eou("اممممم"))        # LOW  - incomplete
print(predict_eou("أبي أدفع و"))    # LOW  - trailing conjunction

from livekit.plugins.turn_detector.multilingual import MultilingualModel
from livekit.agents import AgentSession
import livekit.plugins.turn_detector.models as td_models

td_models.HG_MODEL = "uselevers/livekit-turn-detector-arabic"

session = AgentSession(
    turn_detection=MultilingualModel(),
)