Streaming speech-to-text pipeline that transcribes live audio and masks toxic spans on the fly. Two-stage classifier keeps cost low on clean input, a token-level tagger redacts only the offending words, and the fast path stays under 400ms end-to-end on CPU.

