Reproduce “Low-Latency LLM-Based Streaming Multi-Talker ASR with Verifiable Rewards” 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

Jing Pan, Ruizhi Li, Ju Lin, Jing Zheng, Florian Metze

The authors post-train a streaming speech model for wearable, two-person conversations, rewarding both fixed-role word accuracy and prompt turn starts. Speaker-attributed word error rate (SA-WER) counts word errors and words assigned to the wrong persistent speaker role. On 1,181 scripted recordings across eight language-pair directions, the combined reward improved SA-WER and onset delay; premature emissions increased.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.