How many calls can you make at the same time with a PBX? Approximately how much does it cost? Assume we have fiber internet if that matters
It seems very doable to waste 5-15 minutes with an ai bot, especially given that we know their scripts. And I’ve seen kitboga videos where he’s doing this with his ai. So I was curious how much it would cost to do this at scale, and how big that scale would be
Doing some research, looks like about 25 concurrent calls is the biggest pbx I found on Amazon. The most i found was 100 concurrent calls for $1500 on Google.
It depends on the minutes, most carriers charge you per min
I am an integration consultant for a certain popular phone system that is basically a glorified web UI ontop of a certain PBX server that is slowly branching into automating a significant amount of IT related needs for end users. At a base, it depends on usage/etc, but realistically a PBX server on modern hardware doing basic calls can handle as many calls as your bandwidth can handle. This however would be illegal, so lets not do that.
When you start adding fancy stuff like voice recongition/etc is where you encounter problems. State of the art voice recognition does not run well on CPUs, it’s nowhere near as fast as real time. You need GPUs, for commercial deployment usually you want expensive nvidia proprietary cards with ECC memory/etc. My shitty at home pbx test box is an i3 10105 with a pathetic $80 nvidia 2GB 1030. I have voice recognition embedded into the callflow, so I’m able to shove call audio into the GPU and get it transcribed at about 30x realtime. 20 seconds of audio takes less than 1 second to process. Assuming that the bot is talking half the time, that means that I could sustain 40 calls at once and be fairly responsive. Downside is that GPU memory allocation is a bitch, and moving data in and out causes some delays. Trying to put too much data in the GPU at once will cause a crash. I have some multithreaded locking in place, but that’s still not 100%.
Assuming the rest of the tech matures. I do have a mining motherboard I never bothered using that can handle 8 GPUs at once. Assuming transcription runs at 20x realtime. Every second, it could handle a theoretical 20 seconds * 8 seconds worth of audio. Assuming the bot is talking half the time (and not running voice recongition on itself), that would mean a theoretical peak of 320 calls at once. Of course those are theoretical maximums and assuming GPUs instantly free up memory when done, for real world situations you’d probably want to use at most one third of that to avoid issues.
Your last 2 posts are exactly what I was looking for, thank you.
So the hardware costs seem pretty reasonable, but doing some quick math on the per minute charge looks pretty rough.
0.5c per minute × 60 minutes per hour x 10 hour shift x 25 days per month = ~$75/mo per line
Using voip.ms prices. Are there any carriers that offer unlimited minutes?
Are you currently waiting for scammers to call or placing outbound calls too? Any legal concerns with mass calling?
Are you using python? I’d be happy to discuss/look at your code. I have some thoughts on how you could maybe identify the type of scam and use knowledge of the basic scripts to help the bot
Would be really cool to call your bot, I’ll dm you for the number
So, Voip.ms is not the cheapest by far. There are carriers that offer “unlimited” but usually they cut you off at a certain amount or cost more than flat rates. When you’re talking about multiple lines running concurrently for a whole month, usually you start to qualify for some really cheap rates.
Most of the calls are inbound, with exception of companies that have utilized robodailers to call me. I have the capability for outbound calling, but do not want to do fully automated outbound calling to certain numbers for certain legal reasons. Automated business to business calls on non-tollfree numbers are generally permitted. Toll-free numbers are slightly trickier.
As for source code, 100% will not be sharing it. Sorry. This could eventually be industry changing, with AI replacing the need for physical scammers. It currently has the capability to remember states (identify call types)/etc. I plan on sitting down an working on a giant call tree for different types, part of the problem right now is most of the responses are in the “root” of the call tree…so it will pick responses for tech support when on a “special promotion” call. It’s more just a matter of time and effort. Right now, the next big step will probably be to work on dynamically generating names and addresses, then I’ll start mapping out call flows.