A Kigali radio station captions its call-in show live with a Whisper-shaped model, re-reading the last 30 seconds every second. Follow 30 seconds of audio through the log-mel image, the convolutional stem, the encoder and the decoder, with the shape at every stage and the time it takes. Then feed in audio at the wrong sample rate, remove the stem, or use the full-depth decoder. This lesson needs JavaScript.