DuplexJev enables high-speed speech decisions without decoding
October 1, 2026
DuplexJev uses a small connector to feed ASR-encoder hidden states into a frozen LLM, enabling 80 decisions across eight utterances in 0.1s on an 8-GPU node. By bypassing autoregressive decoding, the system maintains high accuracy in spoken QA and achieves 90% accuracy for speaker gender and emotion.
HOW THIS AFFECTS YOU
●
builderYou can achieve much lower latency for full-duplex voice agents by avoiding decoding steps.
●
researcherThe method shows how single-token supervision on frozen LLMs can capture paralinguistic features like emotion.