Waqf-AI/egyptian-arabic-eot-qwen35-0.8b

Egyptian Arabic End-of-Turn Detector — Qwen3.5-0.8B

A fine-tuned Egyptian Arabic end-of-turn (EOT) detector for voice-agent transcripts. Given what a user has said so far and optional preceding agent context, it estimates whether the agent should respond now or keep listening.

The problem

Voice activity detection can detect silence, but silence alone does not reveal whether someone has finished speaking. In Egyptian Arabic, a speaker may pause while they are still dictating a number, revising a phrase, or about to complete a request. An EOT detector adds the semantic decision:

• —EOT — the turn is complete; the agent may answer.
• —ONGOING — the user is still speaking; the agent should wait.

This checkpoint is intended as a text-based semantic signal alongside an audio VAD, not as a replacement for VAD.

Training data

Fine-tuned on `Waqf-AI/egyptian-arabic-turn-detection`:

The dataset is Egyptian Arabic (arz), context-aware, and split by dialogue_id so truncated variants of the same utterance do not leak between training and test. It is synthetic and author-labeled; results therefore measure performance on this task distribution, not a guarantee of real-call performance.

Method

Base model: `Qwen/Qwen3.5-0.8B`.

Two decision tokens were added and learned with causal-LM supervision:

• —<|turn_done|> — EOT
• —<|turn_continue|> — ONGOING

At inference, the score is the normalized probability of <|turn_done|> versus those two decision tokens. Training used full fine-tuning for 3 epochs, learning rate 2e-5, batch size 4, gradient accumulation 8, and a maximum sequence length of 512.

Held-out results

Evaluated once on the dataset's grouped test split (2,516 rows; 1,267 EOT, 1,249 ONGOING):

The selected equal-cost operating threshold is 0.51. In a voice agent, false EOT causes an interruption and is usually more costly than a late response; tune the threshold on your own validation calls accordingly.

Examples

With preceding agent context في اي خدمه (“How can I help?”):

Minimal inference

import torch
from transformers import AutoModelForMultimodalLM, AutoTokenizer

MODEL_ID = "Waqf-AI/egyptian-arabic-eot-qwen35-0.8b"
DONE, CONTINUE = "<|turn_done|>", "<|turn_continue|>"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, padding_side="left")
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID, dtype=torch.bfloat16, device_map="cuda"
).eval()

context = "agent: في اي خدمه"
text = "تعالي امم انا كنت عايز اسال عن الفاتوره"
prompt = (
    "حدد هل المتحدث أنهى دوره أم سيكمل الكلام. أجب برمز واحد فقط: "
    f"{DONE} أو {CONTINUE}.\n\nسياق المحادثة:\n{context}"
    f"\n\nكلام المستخدم حتى الآن:\n{text}\nالقرار:"
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.inference_mode():
    logits = model(**inputs).logits[:, -1, :]
ids = tokenizer.convert_tokens_to_ids([CONTINUE, DONE])
p_eot = torch.softmax(logits[:, ids].float(), dim=-1)[0, 1].item()
print(f"p(EOT) = {p_eot:.3f}")

Limitations and responsible use

• —This is a text-only model. It does not hear prosody, interruptions, or non-speech audio.
• —It was trained on synthetic, author-labeled Egyptian Arabic examples. Validate on consented, representative production transcripts before deployment.
• —Use a conservative threshold and monitor false EOT decisions: an interruption is typically more harmful than a small response delay.
• —It is not intended for high-stakes decisions or as the sole control for a safety-critical system.