DuplexJev Cuts Autoregressive Decoding for Voice Agents, Answering 80 Speech Decisions in 0.1s

adventists-ai · hf · 2026-10-15

• Researchers from Adventists.ai and Zhejiang University propose DuplexJev for the many small closed decisions in full-duplex voice agents: ASR-encoder hidden states flow through a small connector into a frozen LLM, and each question is read as a single-token distribution over its options—no autoregressive decoding at all.
• An 8-GPU node answers 80 decisions about 8 utterances in 0.1s. With a last-layer connector, spoken QA stays close to reading the transcript (90% vs 91%).
• A cross-attention connector lets the model hear the speaker: gender and emotion accuracy both hit 90% (from 55% and 28%), at just 1 point of spoken QA cost (83%→82%).
• Key training choice: decisions are trained with cross-entropy on the read-out answer token instead of the usual transcript distillation, whose teacher never hears the voice; distillation is kept for content.
• Encoders and LLMs are interchangeable; weights, training recipe, a batched-inference pipeline for full-duplex serving, and a bilingual spoken-QA set are released.

Original post →

More from Research

• NVIDIA report: agentic setup lets GPT-6 Astra hit 74.8% on RoboLab-120 robot manipulation — YuXiang_IRVL · 2026-10-15
• Bittensor Subnet 55 completes 10M synthetic CRISPR experiments with 248 competing miners — markjeffrey · 2026-10-15
• Logit lens shows Qwen3-1.7B's refusal to be shut down depends on how many layers you train — Sauers_ · 2026-10-15
• DeepSeek's first LLM paper was literally titled "Scaling Open-Source Language Models with Longtermism" — teortaxesTex · 2026-10-15
• Vals AI Puts Frontier Models in Real-Time MuJoCo Robot Simulations — scaling01 · 2026-10-15
• AI Post-Processing Tool Extends Extreme Weather Warnings From 10 Days to 6 Weeks — LesterMackey · 2026-10-15

Research channel →