insanely-fast-whisper

This is a genuinely useful project: it takes Whisper (someone else's model) but applies concrete optimization techniques (flash-attention-2, fp16, batching, better-transformer) that produce measurable speedups. The 98-second claim for 150 minutes on A100 checks against the benchmarks shown. No snake oil—just pragmatic engineering. The CLI is opinionated by design, not by laziness. Weak points: not novel research (applying known optimizations), and the README oversells with 'insanely fast' whe...