@AI Evy Hey, we're running a German-language voice agent (GPT Realtime mode, temperature 0.00) for a real estate company. The agent collects email addresses by asking callers to spell them out letter by letter. The problem: When a caller spells a name like G-E-H-R-I-K-E, the agent consistently auto-corrects it to "Gehricke" — changing the single "k" to "ck" because that's a more common German spelling pattern. The caller spells it correctly, the agent even reads it back correctly during the call, but when it passes the value to a function call (e.g. to send an email), it writes "ck" instead of "k". This happens specifically with k → ck, but we suspect it's a broader issue where the Realtime model applies German spelling rules to letter-by-letter input instead of treating it as a raw character sequence. What we've already tried: • LLM Temperature set to 0.00 • Explicit prompt instructions to NOT auto-correct spelled content • Letter-by-letter readback with confirmation • The agent hears and repeats the correct spelling — but still outputs the "corrected" version Since we're on GPT Realtime (speech-to-speech), we don't have access to boosted_keywords or custom STT config. Is there any known workaround for this in Realtime mode? Has anyone found a way to force the model to treat spelled-out sequences as verbatim input without applying language-model corrections? Thanks!

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