Realtime API does not account for user corrections

API
1

Hey all,

I spotted some weird behavior from the real-time API, which I recreated through the playground (see image).

Basically, what I’ve noticed is that when the user tries to make a correction to the way the model initially interpreted something, the model just ignores this correction. In the screenshot, for example, the model ignores me correcting the spelling of my name (it still spells it wrong after my correction).

Has anyone else noticed this sort of strange behavior? Anyone know of a fix?

Thanks!

This image shows an interface with a simulated chat session between an assistant and a user named Priscilla, who is asked to provide personal information, while system settings for voice activity and model details are displayed on the right. (Captioned by AI)
This image shows an interface with a simulated chat session between an assistant and a user named Priscilla, who is asked to provide personal information, while system settings for voice activity and model details are displayed on the right. (Captioned by AI)1641×922 84.8 KB

1 Like
2

i have faced the same issue. My use cases are on date/time correction.