Can anyone familiar with the technical details shed light on why this is so.

Is it because of a globally trained model (as opposed to trained[tweaked on] on context specific data) or because of using different classes of models.

Neither copilot nor flow can natively handle audio to my understanding, so there is already a transcription model converting it to text that then GPT tries to summarise.

It could be they simply use a mediocre transcription model. Wispr is amazing but would hurt their pride to use a competitor.

But i feel it's more likley the experience is; GPT didn't actually improve on the raw transcription, just made it worse. Especially as any miss-transcipted words may trip it up and make it misunderstand while making the summary.

if i can choose between a potentially confused and misunderstood summary, and a badly spellchecked (flipped words) raw transcription, i would trust the latter.