This guy in The New Yorker thinks highly of it: https://www.newyorker.com/tech/annal...modular-future .

I haven't been able to get it to run yet. I get
https://github.com/ggerganov/whisper.cpp

Later:

I got it to work: one needs to download a database first to build successfully. I also can't make the '-m' switch, which points to the directory of the databases, to work, but run in its directory.

It works well: it gets many rare words, punctuates mostly correctly, which means it gets the ends of sentences correct. It identifies quotes, usually correctly. It gets most proper nouns (i.e., capitalized) correct.

Very interesting. Runs on a RazPi, huh? Two questions:

1. Where do you get the database?
2. Do you know how big a chunk of speech it will take at a time?

One other thing: is it half or full duplex? Will it translate and listen at the same time?

It seems like this only makes sense to use on Macs?

https://github.com/ggerganov/whisper.cpp/issues/430

From the same place I got the source. I didn't read the README before I built the first time.

I've used it on hour-long files. It takes 4 hours on my 14-year-old single CPU 4 processor computer for an hour's worth of speech.

Half.

I use it on not-Mac.

Will it work from a list of files or would a person have to script it to do so? Thanks for posting this, it sounds very interesting.

Thanks for sharing. This looks very interesting.

I don't know. How's thor?

Unworthy.

After 20 transcriptions of hour-long programs that took about 4 hours each, the 21st took 10, the 22nd took 6 hours to transcribe 5 minutes before I killed it. I downloaded the latest and it seems to be working as well as it did before. Odd.

4 hours working on 1 hour of speech? I doubt if I'll try it on a RazPi, then.

I got ffmpeg equipped on the Razpi. But whereas my box re-encodes 1080p videos at about 3.75x, the RazPi only achieved 0.2x. So it would take 5 hours to re-encode a 1 hour video.

It accepts multiple input files.
I have a 14-year-old laptop. I hope a modern Pi is more powerful. It won't hurt to try.

Nope, the Pi will hardly be more powerful

A twin core laptop will have two threads per core, and so have a maximum of 400% in top. The Pi is 4 core but only does 1 thread per core. Your laptop will be clocked faster. The cores sare also more powerful.

Arm charges a premium for it's beefy cores so people are buying A-72 & A-76 cores because you can bundle those into price sensitive SoCs. They are inclined to be fabricated in klunky (by today's standard) wafer fab lines. So current & wattage increases with speed. There's half a motherboard in those SoCs.

You can use a smaller model. I use the base. You can try the tiny model, which is half the size.

I found a glitch: it takes huge amounts of time on singing that is close to intelligibility. Unintelligible singing it codes as music without taking a lot of time; clear singing it transcribes about as well as talking. It has a switch, -ot , that sets the point in the file at which it starts transcribing. I kill a process hung up on singing and restart it after the singing.

@RandomTroll - Thanks for the on-going review.