We're seeing inconsistent behavior when our agent reads back phone numbers to callers. Sometimes it reads digits naturally (e.g., "four four five, five five five, one two three four"), other times it reads them in awkward groupings or rushes through without pauses. This happens across different calls with the same prompt instructions. We've tried formatting the number various ways in the prompt (with dashes, parentheses, spelled out) but results are still unpredictable. Anyone found a reliable way to get consistent phone number pronunciation? Or is this something on the TTS/model side that needs addressing?

Hi @AkivaG

call_af11a0c8e0ab4ca252ac9f0e0fa GPT 5.2

Can you use GPT 4.1 as it has more accuracy and is recommend