Overview
This model, Scicom-intl/Malaysian-Turn-Detector-Qwen3-1.7B, is a specialized fine-tuned version of the Qwen3-1.7B base model. Developed by Scicom-intl, its primary function is real-time turn-end detection in multilingual call center environments. It operates by predicting the probability of the <|im_end|> token, indicating whether a speaker has finished their turn, making it suitable for low-latency voice agent systems.
Key Capabilities
- Precise Turn-End Detection: Achieves an overall accuracy of 96.67% with a 0.5 probability threshold, demonstrating high precision (99.82%) and recall (93.50%).
- Multilingual Performance: Evaluated across various language pairs including Chinese, English, Malay, and Tamil, showing robust performance with per-language accuracies ranging from 94.00% to 100.00%.
- Optimized for Voice Agents: Designed to integrate into pipelines like LiveKit, enabling voice agents to respond promptly and naturally.
Training Details
The model was trained on positive samples (complete conversations) using a Liger Fused Linear Cross Entropy loss function, Flash Attention 3, and bfloat16 precision. The training dataset includes diverse sources such as Call Center Language Switching, Function Call, Malaysian Multiturn Chat Assistant, and Malaysian Speech Instructions datasets.