Hi @AI Evy Good day Hi Retell Support, We’ve noticed inconsistent STT behavior during phone-number collection. When a caller first provides their complete phone number, STT(Cartesia) sometimes transcribes it incorrectly. However, after the agent asks the caller to repeat the same number, STT often recognizes it correctly on the second attempt. We’ve observed this particularly with repeated digits and compound spoken number groups, including: 347-444-4844 929-929-8555 518-331-0956 855-748-5596 We also confirmed a specific example in a call: At 1:52, the agent said: “I heard the phone number as 855-748-5596. Is that correct?” At 2:00, the caller corrected it: “No. It’s eight four five, seven four eight, five five nine six.” We reviewed the call audio, and the caller had also said “eight four five” the first time. However, the first recognition changed 845 to 855. When the caller repeated the number, it was recognized correctly. Is this expected STT behavior when handling phone numbers, particularly repeated digits or grouped/compound spoken numbers? Is there an STT setting, numeric recognition option, or recommended native phone-number intake method that can improve first-attempt recognition accuracy? Also, can Retell expose word-level confidence scores or raw STT output so we can distinguish an STT recognition error from downstream number extraction? We’ve already added Extract Dynamic Variables and Code nodes for validation. However, these can only validate the STT output—they cannot recover digits that STT mistranscribes on the first attempt. We’d appreciate any recommended configuration or best practices for improving first-attempt phone-number recognition accuracy. Thank you!

Thanks for tagging me @Gene. I'll look through relevant sources and get back to you shortly.