研究

Simultaneous speech translation has to produce output before the speaker finishes. Commit too early and the translation bets on the wrong future; wait too long and latency defeats the purpose.

The idea: instead of one guess about how the sentence ends, consider many plausible continuations and only emit the target tokens that hold up across them. The training data for this behavior is synthesized without human reference translations.

Method details, numbers, and figures will appear here once the work is public.

IO-aware fusion, Tensor Core pipelining, and per-workload autotuning applied to sparse attention for long-context inference. See the project page for the breakdown.