In our case, we don't/can't analyze the audio stream of consumer calls. We can only work with the meta data of the call.
Our algo comes down pretty hard and fast on calls that come into our honeypot. These numbers have been out of service for years and no one should be calling them. Since these phone lines belong to us, we can do anything we want with them - there's no privacy concerns like there would be with consumer calls.
We answer, record, transcribe, analyze, and classify almost 10k calls per hour coming into our honeypot. Check out https://www.nomorobo.com/lookup. There's been a ton of health insurance scams running this week but the scam du jour changes every day.
When we analyze the call data coming into consumer lines, we have to be a little less aggressive so that we don't accidentally block schools, police, doctors, pharmacies.
The only thing left for humans is to help correct the decisions the algo makes. Sometimes it misses robocalls and sometimes it stops things it shouldn't. People just report them through our app and website.
For historic analysis, we also ingest the FTC & FCC robocall complaint data sources. But it's not really a great data source for the real-time detection algo - It's just too slow to be actionable.
You really have to be detecting robocalls at the moment they come in to have a successful blocking product.
But, the response to our sampling system has been really strong so we're expanding that to grab one recording per day per number. That way you can also go back in time and see the different scams that have been sent from the same number.
We're also building some deeper analytics so that you can see call volume by robocaller and graphs that show the changes over time.
There's a business called Pindrop that conduct an analysis of audio data to guess where a call is originating from. The idea being, if a call originates from somewhere like India it'll take roughly a certain route that will produce certain audible artifacts on the line.
They also score a bunch of other stuff. I'd recommend reading the patents their CEO filed if you're interested. It's interesting.
I wondered if you could share, roughly, what your algorithm is doing? Are you also analyzing audio data?