JAL-Turn: Joint Acoustic-Linguistic Modeling for Real-Time and Robust Turn-Taking Detection in Full-Duplex Spoken Dialogue Systems

Dimensions overall score 7.0

• JAL-Turn is a lightweight and efficient speech-only turn-taking framework that adopts a joint acoustic-linguistic modeling paradigm.method 95%
  Evidencepartial

  In this paper, we propose JAL-Turn, a lightweight and efficient speech-only turn-taking framework that adopts a joint acoustic–linguistic modeling paradigm

  Implicationpartial

  This is a core claim stated directly in the abstract and elaborated in the introduction.

  Verificationpartial

  partial

• JAL-Turn enables turn-taking prediction to run fully in parallel with speech recognition by sharing a frozen ASR encoder, introducing no additional end-to-end latency or computational overhead.technical 95%
  Evidencepartial

  By sharing a frozen ASR encoder, JAL-Turn enables turn-taking prediction to run fully in parallel with speech recognition, introducing no additional end-to-end latency or computational overhead.

  Implicationpartial

  This technical advantage is explicitly stated in the abstract and further detailed in the method section.

  Verificationpartial

  partial

• JAL-Turn consistently outperforms strong state-of-the-art baselines in detection accuracy while maintaining superior real-time performance on public multilingual benchmarks and an in-house Japanese customer-service dataset.result 95%
  Evidencepartial

  Extensive experiments on public multilingual benchmarks and an in-house Japanese customer-service dataset show that JAL-Turn consistently outperforms strong state-of-the-art baselines in detection accuracy while maintaining superior real-time performance.

  Implicationpartial

  This is a primary result claim, directly stated in the abstract and supported by multiple tables in the results section.

  Verificationpartial

  partial

• JAL-Turn achieves 92.03% accuracy and 0.925 F1 on an in-house Japanese corpus, representing significant gains over STurn-v2.result 90%
  Evidencepartial

  JAL-Turn92.03 0.92538

  Implicationpartial

  Specific numerical results are provided for accuracy and F1-score on a particular dataset, with a direct comparison to a baseline.

  Verificationpartial

  partial

• JAL-Turn achieves latencies of 22 ms on a public dataset and 43 ms on an in-house corpus, operating within the real-time regime.result 90%
  Evidencepartial

  JAL-Turn operates comfortably within the real-time regime, with end-to-end latencies of 22 ms on the public dataset and 43 ms on the in-house corpus.

  Implicationpartial

  Specific latency figures are provided for different datasets, demonstrating real-time performance.

  Verificationpartial

  partial

• JAL-Turn underperforms on the backchannel (bc) state compared to some baselines due to the context-dependent nature of backchannels.limitation 85%
  Evidencepartial

  JAL-Turn underperforms on the bc state (80% vs. 91%), which we conjecture stems from the intrinsically context-dependent nature of backchannels

  Implicationpartial

  This is a specific limitation identified and explained in the results section.

  Verificationpartial

  partial

• A scalable data construction pipeline is introduced that automatically derives reliable turn-taking labels from large-scale real-world dialogue corpora without manual annotation.method 95%
  Evidencepartial

  We introduce a scalable data construction pipeline that automatically derives reliable turn-taking labels from large-scale real-world dialogue corpora without manual annotation

  Implicationpartial

  This is a key methodological contribution explicitly stated in the abstract and bullet points.

  Verificationpartial

  partial

Competitive landscape

A lightweight, speech-only framework for real-time, robust turn-taking detection in voice AI agents, integrating acoustic and linguistic cues without adding latency.