Now, not only does someone not need to be actively listening, they don't even need to save the audio. Just convert it to text, keep it indexed and whatever three letter agency is doing the tapping has an easily searchable record of your communications should they ever feel the need to look into you without you knowing.

Converting to text loses data and will fail on unexpected languages or accents. It's better to use a specialized voice codec, which can have fantastically high compression rates, or just keep the (already annoyingly-companded) 8kbit voice stream around in its entirety. It's pretty small.

You must mean 8Khz - which results in a 64kbps stream. (8000 bytes per second) The companding was actually a very good use of 8 bits per sample for voice, introducing little artifacts except at high amplitudes and the low pass filter. Nowadays I find it ridiculous mobile networks feel it's still necessary to compress the audio further - 64Kbps is nothing on modern mobile networks, ie VoLTE etc... WB-AMR is definitely an improvement with it's 16Khz sample rate and a bitrate lower than that of G.711, but mostly not supported between different mobile carriers...

You convert it to text in order to index it. If it becomes interesting in the future, you listen to the audio (maybe after getting permission from a judge in a secret court.).

Recording all calls is well within the financial and material means of the US Government. http://blog.archive.org/2013/06/15/cost-to-store-all-us-phon...

I agree. The convert to text part was about indexing the content of the calls for easy searching.

I bring it up because speech recognition has become so commoditized that most of us could think of a way to whip up an, albeit bad, solution to this problem using AWS/GCP/etc in a weekend.

Many years ago, I worked for a guy who regularly monitored mobile phone frequencies and recorded conversations. He wrote screenplays and used the recordings for dialog study. At least that was what he claimed.