Reproduce “Batched Speech Decisions Without Decoding: Single-Token Supervision Lets a Frozen LLM Hear Beyond the Transcript” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jie Jin, Ziyin Ma, Min Yin, Jinyu Chen, Haigang Song, Zhikun Pang, Xiaowen Zhang
DuplexJev reads answer choices from a frozen language model’s next-token probabilities, using speech-encoder features and a trainable connector rather than generating text. It batches decisions in one pass. Answer-token supervision lets tested models use voice cues that transcript-only distillation misses. In separate settings, content reaches 90/100, while mixed-training gender and emotion accuracy reaches 89.9% and 90.0%, respectively; however, ZJU-ML drops 16 points.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.