Reproduce “Factorized Delayed Streams Modeling for LLM-based Streaming ASR” 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
Tatsunari Takagi, Kai Washizaki, Atsushi Kojima, Lianbo Liu, Koki Nikaido, Yui Sudo
Factorized DSM (F-DSM) adapts Delayed Streams Modeling for speech recognition by separating waiting from text-token choice. A small head predicts whether to wait, while the original LLM vocabulary head predicts words when ready. At 480 ms delay it reports 6.17% average LibriSpeech WER versus 6.61% for DSM without <w> on the tested Qwen3.5-0.8B system and datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.