Batched Speech Decisions Without Decoding: Single-Token Supervision Lets a Frozen LLM Hear Beyond the Transcript

2026-10-01

AI Summary

This work addresses the high response latency caused by conventional autoregressive decoding in full-duplex voice agents. We propose DuplexJev, an architecture that feeds hidden states from an ASR encoder into a frozen large language model via a cross-attention connector. By employing a single-token classification mechanism to directly read decisions, it enables decoding-free, ultra-fast batched inference. Furthermore, we replace transcription distillation with cross-entropy supervision on answer tokens, allowing the model to perceive gender and emotion information beyond textual content. Experimental results demonstrate that on an eight-GPU node, DuplexJev completes eighty decisions within 0.1 seconds, achieving nearly 90% accuracy across spoken question answering, gender recognition, and emotion recognition tasks.

Abstract

Full-duplex voice agents make many small, closed decisions, which current systems answer by slow autoregressive decoding. We propose DuplexJev, which feeds ASR-encoder hidden states through a small connector into a frozen LLM and reads each question as a single-token distribution over its options. Nothing is decoded, and an 8-GPU node answers 80 decisions about eight utterances in about 0.1 s. With a last-layer connector, spoken QA stays close to reading the transcript (90% vs. 91%). DuplexJev also hears the speaker: gender and emotion accuracy both reach 90% (from 55% and 28%) with a cross-attention connector, whose spoken QA drops by only 1 point (83% to 82%). We train decisions with cross-entropy on the read-out answer token, instead of the usual transcript distillation, whose teacher never hears the voice, and keep distillation for content. Encoders and LLMs are interchangeable; we release weights, training recipe, a batched-inference pipeline for full-duplex serving and a bilingual spoken-QA set.

Problem

• full-duplex voice agents
• autoregressive decoding
• spoken QA
• paralinguistic information
• transcript distillation

Innovation

• Single-Token Supervision
• Decoding-Free
• Frozen LLM
• Cross-Attention Connector
• Full-Duplex Voice Agent