Whisper models demonstrate effective in-context learning abilities for speech recognition, improving performance without gradient descent.