Conventional wisdom says dictation accuracy depends on microphone quality, proximity, and a quiet room. But research into systems like Whisper tells a different story. Daniel Herman and Corn discuss how modern end-to-end neural ASR models are surprisingly robust to background noise, whispering, and fast speech — and why the single biggest predictor of accuracy is how well your audio matches the model's training distribution. They explore counterintuitive findings from Johns Hopkins, Carnegie Mellon, and ETH Zurich, including why language mismatch in background conversations can actually help, and how humans and machines co-evolve through computer-directed speech. If you dictate text, this episode will change how you think about your setup.
Episode #675224 — open it directly at myweirdprompts.com/675224