V20 U8 BETA 2: Handle Calls Like a Boss!

[patrickb_de]

“Ah, I could’ve just tried that myself. I tested it with .txt and .docx files. It’s clear that .docx doesn’t work - it probably needs an additional library to read it

 

[KyriacosS_3CX]

Ahhh okay. Ive tryed a txt file and a docx. So only pdfs are supported. Okay. Nice to know ^^

No, no! A lot of file types are. I used a PDF as they tend to be larger and so an upload fail was more likely. TXT, CSV, docx, xlsx, quite a few more.

From the Blog post:

Administrators can build dedicated Knowledge Bases and upload a wide range of file formats txt, md, pdf, docx, csv, json, html, pptx, xlsx and more.

 

[patrickb_de]

Hi, could you share one of the files you are trying to upload? Just tried a 5.89MB pdf with success. From public source:
https://opengeology.org/historicalgeology/wp-content/uploads/2022/10/Plate-tectonics.pdf

Now it works. Reboot fixed the problem. i.d.k. why.

 

[ArneDery]

Not for the moment. It's been available to all to try for 1,5 years

So the beta 2 update also disabled the 3CX cloud transcription? I noticed it stopped after doing the update.
Switched back to OpenAI again to be able to demo. Not been able to install the self host AI yet.

 

[tronic]

Sorry if it's just me being slow, but just want it to be clear.

So as long as we are using our own OpenAI key, you can add Transcription for voicemail and recording on any version of 3cx? Or if the 3CX is upgraded to ENT, then they can use the 3CX transcription.

Also, for the local transcription server is this only for one 3CX or can this be cloud hosted on the same virtual network and be used by multiple 3CX systems?

 

[KyriacosS_3CX]

Also, for the local transcription server is this only for one 3CX or can this be cloud hosted on the same virtual network and be used by multiple 3CX systems?

Can be hosted anywhere and used for multiple 3CX systems.

Regarding 3CX Transcription, this is available in various ways to different ENT licences.
Using your own OpenAI/Google keys is a more generally available feature.
Your sales contact should be able to provide a more detailed analysis.

 

[Ratsch]

Our testing system unfortunately has a 4SC, so I can't further test the boss secretary.
How could the Boss redirect to his mobile phone number, if he is out of office for example?
Say someone calls him, it gets redirected to his secretary - how can the secretary reach the Boss' mobile phone number?

Some of our customers would like to decide themself whether or not the call from the secretary should go on his phone or mobile phone (number). Can they just change their status to "Away" for example (set to forward internal calls to "My mobile")?
In the first beta this was not possible as every call will be directed to the PWA/Softphone/Telephone, no matter the status settings.

Also in german it is currently called "Chef-Sekretärin" which implies a female secretary. In germany nowadays you tend to name things neutral (either generic masculine neutral or included neutral), so it would either be "Chef-Sekretär" or "Chef-Sekretär/in" (there is no definite rule on included neutral, so it could also be Chef-Sekretär:in or Chef-Sekretär*in - or generic neutral "Chef-Sekretariat" or "Sekretariatsfunktion").

Another question for shared voiceboxes: If we set "Destination if no answer" and "When office is closed/on break/on holiday" to the queue/group voice box and each route needs a different announcement, how can we have a different mailbox for each route (same like the different voiceboxes set in each status of a user) (this was asked/requested by @rlg but seemed unanswered)

​

 

[Pierre.m]

Great update, now we can build awesome solutions using AI.

Testing some things with the new AI agent, i created two of them (receptionist and assistent), assigned some simple PDF files as resources and completed the setup as in the docs.
I can see that the recource files are transferred to OpenAI, to the API seems to be working well, also assigned the right models.

Only thing is, i cannot get any audio.
The call is connected, i can see it running in the panel, assigned to the right AI but no audio at all.
Anybody else dealing with this issue?

Also checked the codecs, PCMU is the only one i allowed to make sure its using the correct one.

 

[KyriacosS_3CX]

@Pierre.m, we haven't seen this. PErhaps you can try restarting the services?

 

[morefon]

Hi, first of all: great job on the local transcription engine! We’re genuinely impressed by how smoothly everything runs on our own infrastructure, and all customers we’ve announced this option to are absolutely thrilled about it. Especially in consulting and legal environments, having transcription and summaries processed locally solves a huge amount of data-protection and compliance concerns. So really: well done. During our testing we noticed one point we wanted to ask about. The speech recognition quality is excellent and very close to OpenAI and Google, which is amazing.

The only difference is the summarization: the summaries from the local engine are much shorter and more generic, while our customers rely heavily on more detailed, structured summaries for their notes. Here is one of our comparison examples:
We recorded a Youtube Video, with 5 minutes lenght and played to both solutions and received following summaries.

Local engine summary: “The employee, Mr. Blümach, was dismissed while he was on sick leave. The lawyer is convinced that the termination is invalid and that they have a good chance of winning in court.”

OpenAI summary: “The dismissal of a technical draftsman is discussed after he was terminated without valid reason and with insufficient notice. The conversation reviews the requirements for filing a wrongful termination claim. It is decided to file a lawsuit due to the incorrectly handled notice period and to pursue a severance payment, as the employee already has a new job prospect. – File a wrongful termination claim within three weeks after receiving the notice. – File the claim with the labor court in Hamburg, as the place of performance is Berlin. – Review applicability of the Employment Protection Act based on company size and length of employment. – Challenge the termination due to the incorrect notice period. – Consider a potential severance payment if the employee does not wish to return to the company.”


Because summarization quality often depends on how the prompt is set up, our main question is whether it’s possible to customize or override the summarization prompt in the local engine. We would love to adjust it so that the output becomes as rich and structured as what we get from the cloud models. Thanks again for your great work — we really enjoy testing the system!

 

[ArneDery]

Because summarization quality often depends on how the prompt is set up, our main question is whether it’s possible to customize or override the summarization prompt in the local engine. We would love to adjust it so that the output becomes as rich and structured as what we get from the cloud models. Thanks again for your great work — we really enjoy testing the system!

Exactly what I was wondering as well with the 3CX transcription:

Hi everyone,


I’m wondering if I’m overlooking something. We’re currently using the 3CX transcription feature and absolutely loving it!

For voicemails, there’s a parameter called VMAIL_TRANSCRIBE_LANGUAGE that lets us set the transcription language.
This is fantastic because it ensures the agent receives the voicemail transcription in their native language instead of a foreign one.

However, when it comes to call recordings, I haven’t found a similar parameter or prompt to define the transcription language.


Question:
Is there currently a way (or any plans) to add an option for setting the transcription language for recordings?
I know that with OpenAI integrations you can adjust the prompt in the parameters,
but I couldn’t find anything equivalent for 3CX transcription.


Any insights or workarounds would be greatly appreciated!

View attachment 50346

 

[N_G]

Hi, first of all: great job on the local transcription engine! We’re genuinely impressed by how smoothly everything runs on our own infrastructure, and all customers we’ve announced this option to are absolutely thrilled about it. Especially in consulting and legal environments, having transcription and summaries processed locally solves a huge amount of data-protection and compliance concerns. So really: well done. During our testing we noticed one point we wanted to ask about. The speech recognition quality is excellent and very close to OpenAI and Google, which is amazing.

The only difference is the summarization: the summaries from the local engine are much shorter and more generic, while our customers rely heavily on more detailed, structured summaries for their notes. Here is one of our comparison examples:
We recorded a Youtube Video, with 5 minutes lenght and played to both solutions and received following summaries.

Local engine summary: “The employee, Mr. Blümach, was dismissed while he was on sick leave. The lawyer is convinced that the termination is invalid and that they have a good chance of winning in court.”

OpenAI summary: “The dismissal of a technical draftsman is discussed after he was terminated without valid reason and with insufficient notice. The conversation reviews the requirements for filing a wrongful termination claim. It is decided to file a lawsuit due to the incorrectly handled notice period and to pursue a severance payment, as the employee already has a new job prospect. – File a wrongful termination claim within three weeks after receiving the notice. – File the claim with the labor court in Hamburg, as the place of performance is Berlin. – Review applicability of the Employment Protection Act based on company size and length of employment. – Challenge the termination due to the incorrect notice period. – Consider a potential severance payment if the employee does not wish to return to the company.”


Because summarization quality often depends on how the prompt is set up, our main question is whether it’s possible to customize or override the summarization prompt in the local engine. We would love to adjust it so that the output becomes as rich and structured as what we get from the cloud models. Thanks again for your great work — we really enjoy testing the system!

Thank you for your nice comments, appreciated!

However you can not customize the prompt. The engine we are running is a decent model, optimized for transcription. But its not comparable to the latest and greatest models available from OpenAI or Google Gemini which can run different kinds of tasks. You can of course use our transcription and have it summarized by OpenAI or Google ..... This is possible. We cant offer that customization ...... Maybe in a year when the models we use improve....

 

[nikolascx]

Our testing system unfortunately has a 4SC, so I can't further test the boss secretary.

Speak to your account manager to see what options are available.

How could the Boss redirect to his mobile phone number, if he is out of office for example?

He can do this already.

Say someone calls him, it gets redirected to his secretary - how can the secretary reach the Boss' mobile phone number?

The secretary can have a speed dial saved to transfer calls to the mobile.
The secretary can call the boss and the routing rules will send to mobile.

Some of our customers would like to decide themself whether or not the call from the secretary should go on his phone or mobile phone (number). Can they just change their status to "Away" for example (set to forward internal calls to "My mobile")?
In the first beta this was not possible as every call will be directed to the PWA/Softphone/Telephone, no matter the status settings.

I think you should try it. We added exceptions, and tell us what you think of it.

Also in german it is currently called "Chef-Sekretärin" which implies a female secretary. In germany nowadays you tend to name things neutral (either generic masculine neutral or included neutral), so it would either be "Chef-Sekretär" or "Chef-Sekretär/in" (there is no definite rule on included neutral, so it could also be Chef-Sekretär:in or Chef-Sekretär*in - or generic neutral "Chef-Sekretariat" or "Sekretariatsfunktion").

Noted and reported. Thanks for the feedback.

Another question for shared voiceboxes: If we set "Destination if no answer" and "When office is closed/on break/on holiday" to the queue/group voice box and each route needs a different announcement, how can we have a different mailbox for each route (same like the different voiceboxes set in each status of a user) (this was asked/requested by @rlg but seemed unanswered)

Wait for Update 9

 

[nikolascx]

Great update, now we can build awesome solutions using AI.

Testing some things with the new AI agent, i created two of them (receptionist and assistent), assigned some simple PDF files as resources and completed the setup as in the docs.
I can see that the recource files are transferred to OpenAI, to the API seems to be working well, also assigned the right models.

Only thing is, i cannot get any audio.
The call is connected, i can see it running in the panel, assigned to the right AI but no audio at all.
Anybody else dealing with this issue?

Also checked the codecs, PCMU is the only one i allowed to make sure its using the correct one.

@Pierre.m
Thanks for your feedback
When you have no audio first you need to see if Audio in general works. If that works, then you can move to AI.
Assuming audio for non AI calls works, then make a call from a webclient and call the AI Agent

1. Open a chrome tab
2. Go to: chrome://webrtc-internals
3. Place a Call to the AI Agent
4. Inspect the Active PeerConnection -> In webrtc-internals, look for the latest PeerConnection created at call time
5. Expand it and focus on Inbound RTP Audio and Outbound RTP Audio
6. Under Inbound RTP (audio) check:
bytesReceived
packetsReceived
audioLevel

Are the values increasing? Yes or no?

remember that now we (3CX) need to receive Audio from OpenAI. Which means that
Before AI: We needed to ensure that port forwarding is working correctly so we can communicate with VoIP Providers and remote clients
Now with AI: If the above is not working, AI will be broken FOR SURE. So AI needs to sit on a Proper Network configured 3CX.

 

[nikolascx]

Hi, first of all: great job on the local transcription engine! We’re genuinely impressed by how smoothly everything runs on our own infrastructure, and all customers we’ve announced this option to are absolutely thrilled about it. Especially in consulting and legal environments, having transcription and summaries processed locally solves a huge amount of data-protection and compliance concerns. So really: well done. During our testing we noticed one point we wanted to ask about. The speech recognition quality is excellent and very close to OpenAI and Google, which is amazing.

Thank you for your feedback.
Yes we worked a lot to finetune this.
And we are not done yet.

The only difference is the summarization: the summaries from the local engine are much shorter and more generic, while our customers rely heavily on more detailed, structured summaries for their notes. Here is one of our comparison examples:
We recorded a Youtube Video, with 5 minutes lenght and played to both solutions and received following summaries.

We are aware of this.

Local engine summary: “The employee, Mr. Blümach, was dismissed while he was on sick leave. The lawyer is convinced that the termination is invalid and that they have a good chance of winning in court.”

OpenAI summary: “The dismissal of a technical draftsman is discussed after he was terminated without valid reason and with insufficient notice. The conversation reviews the requirements for filing a wrongful termination claim. It is decided to file a lawsuit due to the incorrectly handled notice period and to pursue a severance payment, as the employee already has a new job prospect. – File a wrongful termination claim within three weeks after receiving the notice. – File the claim with the labor court in Hamburg, as the place of performance is Berlin. – Review applicability of the Employment Protection Act based on company size and length of employment. – Challenge the termination due to the incorrect notice period. – Consider a potential severance payment if the employee does not wish to return to the company.”

I understand. There could be more info.
I think the prompt of the summary says "Summarize this in 50-100 words". A simple fix would be to allow AI to make the summary longer like 200 words.

Because summarization quality often depends on how the prompt is set up, our main question is whether it’s possible to customize or override the summarization prompt in the local engine. We would love to adjust it so that the output becomes as rich and structured as what we get from the cloud models. Thanks again for your great work — we really enjoy testing the system!

We default to 50 words (open ai).
You can only modify the prompts for open ai (OPENAI_SUMMARY) and Google (GOOGLE_SUMMARY)
What transcription engine are you using?
Probably you are using our 3CX Transcriber. Here you cannot change the prompt because we control it. But it is set to be the same. This is why its short.

We are discussing internally to allow users to change this. You will benefit if the summary will be 200 words.. Then you will get the content that openAI gave you.

Thanks for the feedback

 

[rlg]

Regarding transcriptions, from what I understand, 8 and 16SC will be able to use 3CX Cloud thanks to the Enterprise PLUS edition.
32SC and above will have to use their own transcription service (so install a local GPU instance for example), right ? And as a result remaining with the Enterprise AI edition will be enough (PLUS edition would not make sense here).

Any thoughts ?
Am I right ?
Thank you again !

 

[ArneDery]

I'm testing the AI assistant agent role, and Looking for some additional information.

Currently when the assistant has received information from a caller, a message is forwarded using chat.
Is there a possibility to have the message be sent by email instead (AI is set to be assistant of a user, email adress is available) of chat, similar as the message we have for call recordings with transcription?

In cases where chat is not enabled on the system. the assistant does not leave a message. Call is ended and info is lost (only to be found in reporting)

 

[triwardhana]

Not see Indonesia Speech language on 3CX Transcritpion engine, would it also provide on next release?

 

[nikolascx]

Not see Indonesia Speech language on 3CX Transcritpion engine, would it also provide on next release?

Halo sayang!!!
try - It should work. Set to Auto Detect. Let me know how it works. Make a recording and check the transcription.
Whilst you are at it, make an AI Agent also and speak to it in Indonesian.
Terima kasih

 

[nikolascx]

I'm testing the AI assistant agent role, and Looking for some additional information.

Currently when the assistant has received information from a caller, a message is forwarded using chat.
Is there a possibility to have the message be sent by email instead (AI is set to be assistant of a user, email adress is available) of chat, similar as the message we have for call recordings with transcription?

In cases where chat is not enabled on the system. the assistant does not leave a message. Call is ended and info is lost (only to be found in reporting)

Thanks for your feedback - Working on it.