OpenAI Voice activity detection (VAD)

[rlg]

Hi,

Could you tell us how is configured OpenAI Voice activity detection (VAD) for AI agents please ?
Do you use server_vad or semantic_vad ?

As per v20u9, and v20u10.

Thank you very much !

 

[rlg]

As per v20u9, Receptionist AI agents can be interrupted too easily, and are very sensible to background noise.
Hence my question.
Thank you again !

 

[BrunoI_3CX]

Hello @rlg

Please note that the internal configuration and parameters used by the 3CX AI Receptionist when establishing an OpenAI Realtime session are not publicly documented.

Additionally, there are currently no user-configurable options within 3CX to manually select the VAD mode (server_vad or semantic_vad), adjust speech detection sensitivity, or modify interruption and turn-detection parameters.

Regarding the reported behavior, could you please confirm which OpenAI Realtime models you are currently using?

Additionally, does the AI Receptionist get interrupted by background noise even when the caller is not speaking, or does this occur specifically when the caller attempts to speak while the AI Receptionist is responding?

Please advise.

 

[rlg]

Hello @BrunoI_3CX
Many thanks for your reply.

Regarding the reported behavior, could you please confirm which OpenAI Realtime models you are currently using?

You're right, this feedback is based on v20u9, so on gpt-realtime-1.5.

Additionally, does the AI Receptionist get interrupted by background noise even when the caller is not speaking, or does this occur specifically when the caller attempts to speak while the AI Receptionist is responding?

Both of them.
The AI Receptionist gets easily interrupted when there's background noise, a suddent brief background noise, or a brief noise from the caller himself ("yes", "ok", "mmmh mmmh" etc...).

Seems like gpt-realtime-2.1, available in v20u10, improves this.
I'll then report back after a few days into production.