Update 4 Final: Affordable Call Transcription with 3CX AI

[bigboy101]

You might have to do a license refresh, what does it show on the dashboard?

I did a license refresh and also restarted the server but that didn't work. Turns out, after an hour or so, I came back to check and all the options are available on the drop down now.

 

[telis]

I have just tried the 3CX transcription feature on 3CX V20 ENT 32SC. I'm only interested in getting the voicemails transcribed for this customer, full call transcription will cause regulatory issues in their industry. My issue is that short voicemails aren't transcribed. If I leave a message like "it's Chris, call me back" OpenAI will transcribe the voicemail and it will appear on the email notification but the 3CX transcription will read "Recording is too short, cannot transcribe."

No one responded to this, so I thought I'd call it out again before I raise it as a new topic or a ticket.

The 3CX voicemail transcription feature seems according to our testing not to transcribe short voicemails that are able to be transcribed successfully by openai.

Most people leave short voicemails.

Can anyone repro this?

 

[NicholasP_3CX]

The threshold for voicemail transcription is 7 seconds.

 

[Lee Cramman]

Found my way here via the voicemail transcription question, but the call transcription and analytics sounds like a cool / useful (revolutionary?) addition. I think this could work really well for us in managing the quality of the interactions in our (smallish 10 desk) call centre, and potentially for other encounters too. I really rate whisper, so if this is anything like as good as that...

Would love to hear feedback from early implementers.

 

[VosSystems]

Love the new recording transcriptions - only one thing, why doesn't it show as a conversation? Now it's just one blob of text and you have to figure out who said what - but it would be better if it was split up into a conversation between Caller and Receiver or something so it's easier to read

 

[LeonidasG_3CX]

Hi there, unfortunately the underlying technology is unable to detect who is who in the transcript at the moment. Future improvements like stereo diarization may eventually allow this to be a thing, but not for the moment.

 

[VoIPTools]

The fact is that 3CX saves recordings in mono, not stereo. But this quickly gets very complicated. For example, what if more than 2 people are speaking? Further, consider if one of the participants in the conversation is a group of people using a conference room phone where all the voices come in to 3CX as one audio stream.

Perhaps at some point in the future AI will be able to separate speakers by their voice attributes, but today we have to make a best effort using the context of the conversation to make an educated guess when the speaker changes. That's what we do with our 3CX Recording Manager. We instruct OpenAI to separate the conversation into paragraphs based on the context of the conversation and it does a reasonably good job. It may not be 100% accurate all the time, but it is definitely an improvement over a single block of text.