OpenAI ?

[kenn10]

In your repository, it looks like support for OpenAI voicemail transcription is in flux. What are the plans to keep this active with more modern versions of ChatGPT? Also, what about speech generation for recordings? I don't want to see support for OpenAI dropped from the product.

 

[ja133]

For VM transcriptions you can just change the OpenAI Transcription Model model to gpt-transcribe

I saw a pull request that starts working on changing the other models that are depreciated but I am sure the developer will chime in on this as well

 

[pbxgeek]

I don't want to see support for OpenAI dropped from the product.

There are no plans to drop OpenAI support for voicemail transcription or speech generation. Could you point me to what in the repository gave you that impression? I’d be happy to clarify.

The current integration works well, and switching to newer models would increase costs by roughly two to five times. With several months remaining before the existing models are retired, there is no immediate benefit to taking on that additional expense.

I’ll push the necessary updates before those retirements take effect to keep these features working. The plan is to maintain support while avoiding unnecessary cost increases for as long as practical.

 

[kenn10]

There are no plans to drop OpenAI support for voicemail transcription or speech generation.

Good. I saw the pull request dropping the older models and wanted to be sure you were keeping support going forward.

 

[pbxgeek]

Well, I jumped the gun yesterday. I started receiving numerous reports that speech generation had stopped working. After looking into it further, we found that the old models created by OpenAI appear to be throttled. This isn’t official, but it’s what we discovered. Generating a 20-second recording with the old model used to take 10 seconds or less; today, it's taking 137 seconds! FS PBX API requests time out after 60 seconds, so generation fails. A new model, which is significantly more expensive, takes only 6 seconds today. We had no choice but to switch to the new models in the latest commit.