Good quality for a phone call is 2 * 64 kbps, but you can go down significantly (probably around 2 * 8 kbps; that's more than GSM half-rate, with modern codecs it could actually sound acceptable). [edit: added spaces to avoid multiplication being misinterpreted as formatting]
More importantly, in order to effectively exploit these phone calls, they'll want to transcribe them. Transcription AI is getting better and cheaper every year, especially when you can use custom hardware. Then, you can store the full voice recording for "interesting" calls (e.g. activists, their contacts, people with "suspicious" patterns etc. - it doesn't have to be accurate, storage is cheap) for a few years, and transcripts for everyone basically forever.
> around 2 * 8 kbps; that's more than GSM half-rate, with modern codecs it could actually sound acceptable
It’s not only acceptable, it’s widely used. I.e. lots of companies use g711 codec with 64kbps for internal intra-office communication and g729 with 8kbps voice bitrate for inter-office communication (with transcoders on the perimeters). It has been quite a standard approach to corporate voip design for years (decades).
That said there is currently a shift to modern high- and adaptive-bitrate codecs, but it’s not instant.
> and g729 with 8kbps voice bitrate for inter-office communication
Why do they do that?
Even if 1000 people are making a phone call at the same time, that's just 64 Mbps each way if go for the full bandwidth, and an office this size surely must have a beefier Internet connection than that. Why skimp on something so cheap?
With a more modern audio codec like Opus, which does better at low bitrates, a minute at phone-line quality would take around 80-120 kB, one third of MP3. Which makes it even cheaper.
This of course doesn't include the cost of intercepting at all relevant points, or the ability to ingest it in real time, or the ability to do any reads on the data (which Glacier doesn't provide real-time).
It is probably still a drop in the federal bucket.
The NSA is said to have 5 zettabytes at their data center in Utah[1]. It’s been rumored they could take a snapshot (I’ve heard 30 days) of the internet and process it there. Not sure if they meant the US internet or everything they consumed combined. All rumors except my one link.
MP3 is about 1MB/minute
Assume 5 minutes per person per day on average
AWS Glacier costs 0.4 cents per GB per month
So I get a yearly cost of:
Which is pretty much $20M/year for the US adult population.