I don't know if this is buggy or if the approach is just flawed, but there are a lot of missed opportunities for clearer mappings here. Another comment showed a

I don't know if this is buggy or if the approach is just flawed, but there are a lot of missed opportunities for clearer mappings here. Another comment showed an example where 'response' maps to 'spahns' which is a correct USian pronunciation of the second syllable. But nowhere in the corpus is the word 'sponse' used - that would have been the natural choice here, surely? Nor is 'spons' used for 'sponsor' or anything else. So it seems like the random walk of the graph is being over-used here, even though the video talks of using substrings as a default. I think focusing purely on phonetics versus edit distance is a mistake and makes this less accessible than it could be.