seligman99

WhisperX along with whisper-diarization, runs at something around 20x of real time on audio with a modern GPU, so for that part, you're looking at around $1 per twenty hours of content to run it on a g5.xlarge, not counting time to build up a node (or around 1/2 that for Spot prices, assuming you're much luckier than I am at getting stable spot instances these days).

You can short circuit that time to build up a node a bit with a prebaked AMI on AWS, but there's still some amount of time before a new node can start running at speed, around 10 minutes in my experience.

I haven't looked at this particular solution yet, but I really find the LLMs to be hit or miss at summarizing transcripts. Sometimes it's impressive, sometimes it's literally "informal conversation between multiple people about various topics"

Along the same lines, in the Windows world:

The current directory is managed with SetCurrentDirectory/GetCurrentDirectory, however the cmd.exe command-line shell also stores the current directory for each drive in an environment variable like "=C:", and the CRT and shell hides all environment variables that start with a "=".

It gets mightily confused if these two concepts of current directory ever diverge.

$0.005 per hour per IP.

Assuming AWS has 50% utilization on IPs they've assigned for EC2, this is a $1.28 billion/yr fee they created.

Scale is fun.

I asked it to summarize the transcript for a podcast about an episode of Star Trek. One minor issue in grammar, but otherwise, it does a remarkable job of making a summary:

> In the Star Trek episode "The Cage", Captain Pike is held captive by a race of advanced televisions. He is offered a life of luxury in a cage, but he realizes that this is not the life he wants and manages to escape. The televisions watch everything on TV and have become too voyeuristic and passive. At the end of the episode, Pike makes the right choice and is able to see the difference between the falsehood of the fantasy and the difficult but more acceptable reality. This episode highlights the importance of exploration and being better than oneself, and is a reminder of the dangers of becoming too passive and voyeuristic. It also shows how technology can be used to create illusions and how these illusions can be used to manipulate people.

Yep, it takes a bit of GPU RAM. I'm using 3 machines with NVidia 3080 or better. I let them go for a few weeks over the winter break when I was mostly disconnected from the tech world. The workers prioritized podcasts I'm personally likely to want to search, and got through almost a third of my archive.

Now it's down to 1 or 2 machines depending on what's going on, so it'll take much longer to finish up, but I'm in no rush.

It's about an hour or two or a day.

This includes data from 1995 on. The early data is backfill of radio shows that transitioned to podcasts and dumped old episodes in their feed at some point. My reader itself started in 2012, I downloaded around 7000 hours of new podcasts, which works out to 1.7 hours per day. So, around 2 hours per day, since I don't listen every day, and to be fair, I haven't listened to every podcast I've downloaded, some don't interest me. But 1-2 hours of listening a day is the sweet spot for me.

Long ago I had my podcast downloader keep all files it downloads and recently I've been using OpenAI's Whisper to go through and create transcripts of the 8000 or so hours of data I have downloaded over the years.

It's very cool to be able to search through and remind myself of something I heard once. Not exactly life changing, but still, nice to be able to quickly drill down and find audio for something when a curiosity strikes me.

Mostly just vanilla JS on that page, and a tiny bit of Python glue code to turn the WebVTT output from Whisper into a data format for the JS.