Fine-tuned TurnGPT Improves End-of-Turn Detection in Turkish Voice Agents
November 20, 2026
A fine-tuned Turkish TurnGPT model outperforms zero-shot LLMs and multilingual detectors in distinguishing turn ends from mid-utterance pauses. Combining text scores with speaker timing features via logistic regression yields the best performance on the Real-TurnTurk corpus.
HOW THIS AFFECTS YOU
●
builderYou can use fine-tuned TurnGPT and timing features to make voice agents feel more natural and less prone to interrupting.