Opus audio encoding in desktop Chromium got 20-23% faster for speech by enabling AVX2 vectorization in libopus. The AVX2 kernels had been present in the codebase since the 1.5.2 roll in April 2024 but were never wired into the build — a 30-line BUILD.gn change fixed that. Benchmarks show speech encode drops 17-23% while music barely moves (~2%), because two of the three AVX2 kernels live in the SILK encoder used for speech. Notably, 96% of the SILK win comes from a single function (silk_NSQ_del_dec), not the one with the biggest per-call gain. The author also highlights how AI assistance (Claude) made the benchmarking and attribution work cheap enough to actually do properly. The fix landed in Chromium 152, and the same vectorization kernels in libopus are available for server-side Opus encoding too.
