Post
5
DuplexJev: a System One model for speech š§
Jev made "System One" decision models the hot idea of the season: state in, typed probabilistic decisions out, many questions in one pass. Text has Kev, Openjev and Jeva; video has VideoJev. Voice agents need it most, yet a text decision model only sees a transcript, and the transcript is where tone disappears.
DuplexJev takes raw audio as its state. A frozen ASR encoder feeds a frozen LLM through a small connector; every question is a choice question read as one token, nothing decoded.
ā” Fast: 10 decisions in 92 ms on one H200 (transcribe + LLM: 1,978 ms), ~170 decisions/s per GPU, under $0.50 per million decisions (4B). A new decision is one line of text.
š Hears what the transcript drops: state of the art in paralinguistic understanding. LISTEN (EACL 2026) + StressTest + MSPB, 1,128 identical questions:
DuplexJev-32B-Para v1.3: 50.9
Gemini 3.8 Flash: 46.9
Qwen3-Omni-30B: 44.6
GPT-audio-1.5: 40.1
Gemini 3.8 Live: 38.9
GPT-realtime-2.1: 34.2
Full LISTEN set: 38.9 vs 33.0 best published.
How it learned to listen: speech LLMs trained by transcript distillation imitate a teacher that never heard a voice (gender ~55%, emotion ~28%). Training the answer token on the correct answer instead takes both to ~90%.
Paper: Batched Speech Decisions Without Decoding: Single-Token Supervision Lets a Frozen LLM Hear Beyond the Transcript (2610.02638)
Models: adventists-ai/duplexjev-6acc69e95e4fc69acd09adaf
Demo: BitKnight/DuplexJev
Code + every model's raw answers: https://github.com/adventists-ai/duplexjev
(Independent project by Adventists AI, Zhejiang University and University of Virginia; not affiliated with TypeSafe.)
Jev made "System One" decision models the hot idea of the season: state in, typed probabilistic decisions out, many questions in one pass. Text has Kev, Openjev and Jeva; video has VideoJev. Voice agents need it most, yet a text decision model only sees a transcript, and the transcript is where tone disappears.
DuplexJev takes raw audio as its state. A frozen ASR encoder feeds a frozen LLM through a small connector; every question is a choice question read as one token, nothing decoded.
ā” Fast: 10 decisions in 92 ms on one H200 (transcribe + LLM: 1,978 ms), ~170 decisions/s per GPU, under $0.50 per million decisions (4B). A new decision is one line of text.
š Hears what the transcript drops: state of the art in paralinguistic understanding. LISTEN (EACL 2026) + StressTest + MSPB, 1,128 identical questions:
DuplexJev-32B-Para v1.3: 50.9
Gemini 3.8 Flash: 46.9
Qwen3-Omni-30B: 44.6
GPT-audio-1.5: 40.1
Gemini 3.8 Live: 38.9
GPT-realtime-2.1: 34.2
Full LISTEN set: 38.9 vs 33.0 best published.
How it learned to listen: speech LLMs trained by transcript distillation imitate a teacher that never heard a voice (gender ~55%, emotion ~28%). Training the answer token on the correct answer instead takes both to ~90%.
Paper: Batched Speech Decisions Without Decoding: Single-Token Supervision Lets a Frozen LLM Hear Beyond the Transcript (2610.02638)
Models: adventists-ai/duplexjev-6acc69e95e4fc69acd09adaf
Demo: BitKnight/DuplexJev
Code + every model's raw answers: https://github.com/adventists-ai/duplexjev
(Independent project by Adventists AI, Zhejiang University and University of Virginia; not affiliated with TypeSafe.)
