My understanding is VAPI specifically does voice formatting to account for this.. Moreover, this is particularly happening for me with gpt-5-mini, didin't happen at all on gpt-4o-* models..