Whisper with tiny.en or base en what is the difference?

[loninappleton]

In the current job I'm running purfview whisper xxx with, first tiny.en which turned out pretty good (to the length of my inspection and before any corrections.) Then that gave me confidence in running base.en the next minimal model in the list which I'm doing now.

In the past some of the models start out ok but in a long job like a 2 hour play they run out of gas.

I was wondering if any presets of number of cores in an old AMD 4 core CPU could increase accuracy but reduce speed. And then how to do that in the Windows environment.

[loninappleton]

I now see in a sample of Task Manager where there is an option to slow below what is set at normal speed in the CPU tab. I don't know if that's a good option. I've had to familiarize myself with a lot of this since not using the commands much. I don't know of any SE speech to text options to run slow for accuracy.

[loninappleton]

Well base.en locked up with about 5 mins to go. I left it along to run but doing another task after an 11 hour run might have glitched it. Then again maybe the old Phenom and an HD wasn't fast enough. I'll first try another run, then move the job to an SSD and my ASUS Ryzen. I'm going to try the run again undisturbed.

[VoodooFX]

tiny.en or base en what is the difference?

tiny is the worst, base is a bit better than the worst.

I was wondering if any presets of number of cores in an old AMD 4 core CPU could increase accuracy but reduce speed.

Use more accurate model: medium, turbo, large-v2 [the best].

[loninappleton]

OK, so slowing down the run speed will not help accuracy or hang-ups.

[VoodooFX]

OK, so slowing down the run speed will not help accuracy or hang-ups.

What do you mean by "slowing down"?
Using bigger model - slows down the run - higher accuracy.

How do you know that it hang-up? Run it directly.

[loninappleton]

On a long run of 11 hours due to using an HD which is too long compared to using SSD, at 5 minutes of the countdown remaining, I quit it because the countdown started to add time and suggested it was hung up. It's possible I had some loose hardware, hard to say.

[VoodooFX]

Look at the Task Manager and check how much free RAM there is when you run it.
To have more free RAM, close all background programs. SE takes memory too, so use Faster-Whisper-XXL directly.

[loninappleton]

I'm stepping through the models from low to high. I think I've gone as far practically I can go at medium-en. But that is further than I thought I could go. I have to error check the completed medium-en sub.

Kudos to Nikse for this new version of SE. It looks really good and gives more detail on these Whisper audio to text jobs. I've made an SE 4.06 portable folder for the desktop as well since I found that I missed the "play" button it. I did ask how that's done in 5.05 here and thanks for that but I have to review it and use more. In the 4.06 it seems easier and more logical.

[loninappleton]

Look at the Task Manager and check how much free RAM there is when you run it.
To have more free RAM, close all background programs. SE takes memory too, so use Faster-Whisper-XXL directly.

Well SE automates a lot of the install stuff needed to run. Will there ever be GUI for Windows? I'm lost at the command line.

[VoodooFX]

SE doesn't automate anything because there's nothing that needs to be installed for it to run.

Use One Click Transcribe tool, it can't be simpler than that.

[loninappleton]

I screen shot the Git hub page and will get the batch file and practice with it. Thanks.

[loninappleton]

I remember having problems finding anything at github, especially what is wanted most... the download. Same today. What is the direct link?
I can't say I can even make that work if it's run at command line. Usually any attempt at that says back: "everything you know is wrong" like an old routine from The Firesign Theatre. Also is this batch file specific to a model or require editing?

[edit] reading the comments at github I see I am not the first to wonder where the download is.

[VoodooFX]

I remember having problems finding anything at github, especially what is wanted most... the download.

Everything is in Releases. You need to press Assets there if it's not the last post.
https://github.com/Purfview/whisper-standalone-win/releases/tag/Faster-Whisper-XXL

Also is this batch file specific to a model or require editing?

You can edit it in Notepad and change the settings you want.
The line with commands by default looks like that:

:: Start processing
"%dp%faster-whisper-xxl.exe" %file_list% -pp -o source --batch_recursive --check_files --standard -f json srt -m medium

Arguments start with "-" or "--", followed by a value for that argument. Some arguments don't take any values, those are called switches.

The "-m" argument is shorthand for "--model", and its value "medium" specifies the selected model size.
You can remove the "-pp" switch to see the transcribed text as the process runs.
You can specify the audio's language using the "-l" or "--language" argument. For example, "-l English". If the language argument isn't set, the language is detected automatically.