Whisper fine-tuned with LoRA for Caribbean broadcast English: 0.0434 word error rate on the private leaderboard, with paired confidence intervals on every comparison, served live from a ZeroGPU Space.
How it knows: Every word error rate is corpus-level with a bootstrap interval, and every comparison is paired on the same held-out clips, which is how a run that looked like progress was shown to be 0.0071 worse.
Nicholas Smith - all work