The point that a transcriber writes the likely spelling, so a near miss on a name simply vanishes, is the clearest explanation of this problem I've read. We run into the same thing with product voiceovers, where the risky words are acronyms more often than surnames. Whether "SQL" is read as letters or as "sequel" is a choice each product makes, and a transcript shows the same three letters either way. Keeping the researched pronunciation and the engine respelling side by side in one lexicon file is the habit I'd push hardest. It turns each fix into something the next video inherits instead of a one off correction. Being upfront that the 32 takes also set the thresholds is a nice touch.