Issues: first of all, transcript quality: those of TED (Chris Anderson's "Technology Entertainment Design" conferences) are curated, those of YT are oftentimes not. Then, discriminating significant and less significant keyframes, but if one only needs "decent" instead of "perfect", good compromises for heuristic algorithms (more "parametrized procedures" or "recipes") can surely be found.
You are tempting me... In an evening one should get something already usable. Better than a presentation, I see the transcript text with aside thumbnails linking to the full-sized images. As someone who sometimes studies material from video, when I am listening or watching I am not in the strict need for text analysis, and when I am working on the text itself I do not need the video. On the other hand, if I think of the MIT OpenCourseware material and similar (Yale etc.), the "blackboard" shots can be precious to have alongside the text.
--
¹ ffmpeg -i video.mkv -vf "select=eq(pict_type\,I)" -vsync vfr -frame_pts true keyframes-%02d.jpg