GHSA-6C4R-FMH3-7RH8 vLLM: Processing differential in multi-channel audio downmixing enables hidden-input/moderation bypass for audio models

Issue Description Librosa defaults to using numpy.mean for mono downmixing tomono, while the international standard ITU-R BS.775-4 specifies a weighted downmixing algorithm. This discrepancy results in: - Inconsistency between audio heard by humans e.g., through headphones/regular speakers and...