3CX OnBoard AI Transcription

[RepairDesk]

We attempted to set up the 3CX OnBoard AI Transcription Engine Server by following the official documentation and would like clarification from the community.

Environment:

• Debian 12 (Bookworm)
• Cloud VM without GPU
• Installation command generated from the 3CX Admin Console

Observed behavior:
When running the installer command:

source <(curl -s https://<pbx-domain>/webmeeting/onboardai/<uuid>)

The installer:

• Starts normally
• Detects Debian correctly
• Checks for NVIDIA GPU using nvidia-smi
• Exits immediately when no GPU is found, with messages such as:
  nvidia-smi: command not found
  No NVidia GPUs VRAM detected

No Docker images are pulled and no services are started. There is no prompt or fallback for CPU-only operation.

Question:
The documentation states that a GPU is recommended, not required, which implies CPU-only deployments should be supported. However, based on this behavior, the installer appears to require a GPU to proceed.

Can anyone confirm:

• Whether CPU-only installation of OnBoard AI is officially supported
• If there is a separate CPU-only installer or undocumented flag
• If anyone has successfully deployed OnBoard AI without a GPU

Looking for clarification or shared experiences to better understand the intended deployment model.

Thank you.

 

[KyriacosS_3CX]

Hello,

An Nvidia RTX 5 series GPU is mandatory.

The document states that the MINIMUM recommended GPU is an RTX card with 24Gb.VRAM.

 

[RepairDesk]

Hello,

An Nvidia RTX 5 series GPU is mandatory.

The document states that the MINIMUM recommended GPU is an RTX card with 24Gb.VRAM.

Thanks for the clarification regarding GPU requirements.

We have the local 3CX Transcription Engine fully installed and running (GPU detected, engine status is “Up”). Currently we have thousands of recordings in the transcription queue, however Max conversions is fixed at 1 and there is no option in the Admin Console to add additional transcription engines or adjust concurrency.

With only one active transcription:

• CPU usage rarely exceeds 2%
• GPU and RAM are largely idle
• The queue continues to grow

Could you please clarify:

• Is single concurrent transcription a hard limitation by design for the local engine?
• Is there any supported way to increase concurrency on the same server?
• If higher throughput is required, is the recommended approach to use cloud transcription providers instead?

We want to make sure we are following the intended and supported architecture.

Thank you.

 

[KyriacosS_3CX]

Hello, could you confirm the GPU model? Normally the Engine should be doing 3 transcriptions in parallel.

 

[N_G]

Its a very long forum post but you have omitted the most important stuff. What GPU, what machine etc. So like this is difficult for us to help you.

 

[RepairDesk]

Hello, could you confirm the GPU model? Normally the Engine should be doing 3 transcriptions in parallel.

I tried this on AWS with G5 instance type, 8GPU and 32GB RAM.
it appears AWS is not supported as transcribe engine was crashing. is it possible to use CPU only mode?

 

[KyriacosS_3CX]

Did you get the instance as a "bare" machine, and then installed Debian 12 and so on, as per the guide, or did it come with another OS/driver package?

Also the 32GB Ram G5 is one GPU / 8vCPUs - not a real problem here, but just mentioning it.

No, there is no CPU only mode.