Hey team,
We're experiencing a critical issue with OpenAI Realtime API's number recognition accuracy in non-English languages. We're currently piloting voice agents in German, Hebrew, and Czech, and encountering significant misinterpretations of spoken numbers:
Examples of misrecognition:
• German: "5 people" → transcribed as "100"; "5" → "50"
• Hebrew: "16" → "12"; "4" → "2"
• Czech: "4" → "2"; "6" → "4"
Why this matters: Numbers often carry critical business context in our use cases. For example:
• Age verification for eligibility checks
• Quantity specifications for bookings/orders
• Time-sensitive scheduling information
Would love to hear from folks who have solved this in some capacity
b
busy-lunch-15701
11/10/2025, 4:43 AM
Not answering your question, but hope this info is useful.
OpenAI Realtime uses separate model for text transcription and stt for LLM input. This means "the transcription text you see" not always equal to "the text LLM got".
c
clever-dinner-59010
11/10/2025, 5:49 AM
I am aware of this the issue is what the llm is getting is way off, is there a way to control the input modality example model only takes text input from a reliable STT provider
b
busy-lunch-15701
11/10/2025, 5:55 AM
I guess you gotta find an alternative model or apply some workaround. On my project, I was tackling on prompt, workflow, and tools instead. For instance, when asking the use for license plate, I will try to query the plate number along with the given car brand or etc. in the system to verify the correct. It's not guarantee but helps a lot.
c
clever-dinner-59010
11/10/2025, 6:07 AM
Yup that eventually happens in the system but the thing is to arrive to a number it’s a frustrating experience for the customer;
For example in my use case i am asking the customer for number of guests for a restaurant use case the model consistently gets the number wrong and then we ask for clarification then again it gets it wrong so this loop eventually becomes frustrating for the end user ( note this is specifically for non english languages )