Batched Speech Decisions Without Decoding: Single-Token Supervision Lets a Frozen LLM Hear Beyond the Transcript
Paper • 2610.02638 • Published
DuplexJev
Speech decisions without decoding: turn-taking, emotion, non-verbal sounds and more from raw speech, in one forward pass.
Note Paper: batched speech decisions without decoding
DuplexJev — speech to decisions
🎧10 voice-agent decisions from raw speech in one pass
Note Try it: ten decisions from your voice (32B)
DuplexJev Vision — voice + picture
👁Judge a voice against a picture in one forward pass
Note Voice + picture together (research preview)
adventists-ai/DuplexJev-32B-Para
Audio-Text-to-Text • 33B • Updated • 175Note Paralinguistics: emotion, non-verbal sounds, style
adventists-ai/DuplexJev-32B-Vision
Any-to-Any • 33B • Updated • 38Note Speech + image (research preview)
adventists-ai/DuplexJev-4B-Para
Audio-Text-to-Text • 4B • Updated • 72Note Small and fast: ~10 GB GPU
adventists-ai/DuplexJev-32B-Stress
Audio-Text-to-Text • 33B • Updated • 10Note Which word is stressed
-
adventists-ai/DuplexJev-4B-Stress
Audio-Text-to-Text • 4B • Updated • 28 -
adventists-ai/DuplexJev-4B-Para-LiveTurn
Audio-Text-to-Text • 4B • Updated • 20