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

Hugging Face Daily Papers Papers
speech-audio full-duplex frozen-llm spoken-qa emotion-recognition inference-efficiency

Summary

DuplexJev 通过小型连接器将 ASR 编码器隐藏状态输入冻结 LLM，将每个多工决策表示为单一 token 分布而无需自回归解码，使 8 GPU 节点在约 0.1 秒内处理 80 个语音决策，同时在口语问答、说话人性别与情感识别上取得高精度。

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.
Original Article
View Cached Full Text

Cached at: 10/15/26, 02:57 AM

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

Source: https://huggingface.co/papers/2610.02638

Abstract

Full-duplexvoiceagentsmakemanysmall,closeddecisions,whichcurrentsystemsanswerbyslowautoregressivedecoding.WeproposeDuplexJev,whichfeedsASR-encoderhiddenstatesthroughasmallconnectorintoafrozenLLMandreadseachquestionasasingle-tokendistributionoveritsoptions.Nothingisdecoded,andan8-GPUnodeanswers80decisionsabouteightutterancesinabout0.1s.Withalast-layerconnector,spokenQAstaysclosetoreadingthetranscript(90%vs.91%).DuplexJevalsohearsthespeaker:genderandemotionaccuracybothreach90%(from55%and28%)withacross-attentionconnector,whosespokenQAdropsbyonly1point(83%to82%).Wetraindecisionswithcross-entropyontheread-outanswertoken,insteadoftheusualtranscriptdistillation,whoseteacherneverhearsthevoice,andkeepdistillationforcontent.EncodersandLLMsareinterchangeable;wereleaseweights,trainingrecipe,abatched-inferencepipelineforfull-duplexservingandabilingualspoken-QAset.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2610\.02638

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper11

#### adventists-ai/DuplexJev-4B Audio-Text-to-Text• 4B• Updated1 day ago • 80 • 1 #### adventists-ai/DuplexJev-32B Audio-Text-to-Text• 33B• Updated1 day ago • 38 #### adventists-ai/DuplexJev-Gemma-31B Audio-Text-to-Text• 31B• Updated1 day ago • 27 #### adventists-ai/DuplexJev-4B-Turn Audio-Text-to-Text• 4B• Updated1 day ago • 40 Browse 11 models citing this paper## Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2610.02638 in a dataset README.md to link it from this page.

Spaces citing this paper3

Collections including this paper1

Similar Articles

Streaming Speech-to-Text Translation with a SpeechLLM

arXiv cs.CL

Presents a SpeechLLM architecture for streaming speech-to-text translation that adaptively decides when to output tokens based on audio, achieving 1-2 second latency with quality close to non-streaming baselines.