A new benchmark, the Full-Duplex Long-Horizon Benchmark (FD-LHB), has been introduced to evaluate the conversational capabilities of full-duplex speech models over extended periods. Research indicates that current models, despite their ability to listen and speak simultaneously, struggle with long conversations, often degrading or exhausting their context windows within minutes. The study identifies two key mechanisms, Initialization-Anchor Eviction and Duplex Repetition Penalty, contributing to model collapse, and proposes simple fixes like Initialization-Anchor Pinning and Duplex Repetition Penalty to mitigate these issues without retraining. AI
IMPACT Highlights limitations in current full-duplex speech models for sustained conversations, potentially guiding future research towards more robust conversational AI.
RANK_REASON The cluster contains a research paper detailing a new benchmark and analysis of AI model capabilities. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Duplex Repetition Penalty
- FD-LHB
- Full-Duplex Long-Horizon Benchmark
- Full-Duplex Speech Models
- Initialization-Anchor Eviction
- Initialization-Anchor Pinning
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
