AIKIDO-2026-545455

TL;DR

The offline speaker diarization pipeline indexes result vectors, an Eigen matrix, and intermediate sample lists using speaker and cluster IDs without validating them against the expected range. When FastClustering produces cluster labels outside the expected count, ReLabel writes past the matrix column bounds, SortBySpeaker reads past the result vector, and the post-filtering index loop reads past the sample lists, causing out-of-bounds memory access. Processing many audio files through diarization can intermittently trigger heap buffer overflow and SIGSEGV crashes of the host process. The fix validates speaker, cluster, and sample indices before array and matrix access and skips out-of-range IDs.