Run noScribe on Linux

[jago-jago]

What a wonderful tool! Thank you very much for this!

I find myself checking quite often these days whether you have found the time to write some instructions on how to make noScribe work on linux. On my main computer I can make it work in a virtual machine (which puts quite a strain on my system) but aside from running it natively I would really love to love to be able to run it over night on a raspberry.

Any idea when you might get around to looking into that? As I have no experience at all with python I am pretty lost trying to figure out how to port this to linux myself.

Again: brilliant work! Thanks!

[kaixxx]

The best instructions I have are these: #33 (comment)
You have to install Python (version 3.9 or bigger) and Git first, if not already on your system. Then follow the instructions above, it should be quite straightforward.
The noScribe Editor will be missing, but you can continue using it in your virtual machine.
I'm pretty sure though that a raspberry pi is to weak for this task...

[jago-jago]

Thank you for this incredibly fast reply! It almost works!

noScribe opens up, but I'm getting the following error message:

Error in options
name 'get_cuda_device_count' is not defined

Do you have any suggestions for that?

But I got it to open up modifying your suggestion as follows (note particularly the underline in "./requirements_linux"):

git clone https://github.com/kaixxx/noScribe.git
cd noScribe
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -r environments/requirements_linux.txt
python noScribe.py

[jago-jago]

Well, now it doesn't even start. I get:

Traceback (most recent call last):
  File "[my-home-folder]/noScribe/noScribe.py", line 39, in <module>
    from subprocess import STARTUPINFO, STARTF_USESHOWWINDOW
ImportError: cannot import name 'STARTUPINFO' from 'subprocess' (/usr/lib/python3.11/subprocess.py)

[gernophil]

Ah, sorry. My bad. It's late, so I missed this :). Please change this (line 37-40):

if platform.system() == 'Windows':
    # import torch.cuda # to check with torch.cuda.is_available()
    from subprocess import STARTUPINFO, STARTF_USESHOWWINDOW
    from ctranslate2 import get_cuda_device_count

to this

if platform.system() == 'Windows':
    # import torch.cuda # to check with torch.cuda.is_available()
    from subprocess import STARTUPINFO, STARTF_USESHOWWINDOW
if platform.system() in ("Windows", "Linux"):
    from ctranslate2 import get_cuda_device_count

[jago-jago]

Okay, interesting.

On my Debian 12 system I get:

Converting audio...
ffmpeg: [mp3 @ 0x759f780] Estimating duration from bitrate, this may be inaccurate


Canceling... (please wait a second)
ffmpeg: [mp3float @ 0x75a20c0] invalid new backstep -1

Audio conversion finished

Speaker identification...
Loading pyannote
Error in step 2 - identifiying the speakers.
cannot access local variable 'diarize_abspath' where it is not associated with a value

And if i try not to identify the speaker it goes on with

Error in step 3 - transcription
Unable to open file 'model.bin' in model '[home folder]/noScribe/models/faster-whisper-large-v2'

And on the raspberry I get

Error in step 1 - converting the audio.
[Errno 8] Exec format error: '[my home folder]/noScribe/ffmpeg-linux-x86_64'

And yes: it is late! Thank you very much for all your work at this hour of night!

[gernophil]

On my Debian 12 system I get:

...
cannot access local variable 'diarize_abspath' where it is not associated with a value

Try to change this

                        if platform.system() == 'Windows':
                            diarize_abspath = os.path.join(app_dir, 'diarize.exe')
                        elif platform.system() == 'Darwin': # = MAC
                            # No check for arm64 or x86_64 necessary, since the correct version will be compiled and bundled
                            diarize_abspath = os.path.join(app_dir, '..', 'MacOS', 'diarize')
                        if not ('diarize_abspath' in globals() or os.path.exists(diarize_abspath)): # Run the compiled version of diarize if it exists, otherwise the python script:
                            diarize_abspath = 'python ' + os.path.join(app_dir, 'diarize.py')

to this

                        diarize_abspath = 'python ' + os.path.join(app_dir, 'diarize.py')

(Remove the original lines 965 to 970 and unintend the original line 971 by one level (meaning remove 4 spaces in the beginning of the line). Remember you added a line before, so the line numbers are different for you. We did not assign diarize_abspath initially for linux. That's something we need to fix on our side, but this solution should be a quick workaround)

And if i try not to identify the speaker it goes on with

Error in step 3 - transcription
Unable to open file 'model.bin' in model '[home folder]/noScribe/models/faster-whisper-large-v2'

That's because you need to download the models separately. Check the manual in models/faster-whisper-large-v2/download.txt and models/faster-whisper-small/download.txt

And on the raspberry I get

Error in step 1 - converting the audio.
[Errno 8] Exec format error: '[my home folder]/noScribe/ffmpeg-linux-x86_64'

The linux executable might not work with the linux distribution you are running on your raspberry. Which Linux did you install there? I guess you need a ffmpeg binary for the distro.

[jago-jago]

You are brilliant!

That's it! Both speaker identification and transcription work nicely - and much faster compared to the virtual machine.

The linux executable might not work with the linux distribution you are running on your raspberry. >Which Linux did you install there? I guess you need a ffmpeg binary for the distro.

I have the standard raspberry os which is a Debian 12 at its core. But you might be on to something with the ffmpeg binary... But since it could not be resolved with a quick search I'll deal with it another day.

[kaixxx]

Nice to see this issue beeing resolved. (I'm on holidays...)
@gernophil: Thanks for all your help! Could you update the main codebase with the necessary changes for linux support?

[gernophil]

Nice to see this issue beeing resolved. (I'm on holidays...)

@gernophil: Thanks for all your help! Could you update the main codebase with the necessary changes for linux support?

I'm on holidays too :D. But I'll return tomorrow and make the changes on the weekend then :).

[gernophil]

@kaixxx:
I think the line

if not ('diarize_abspath' in globals() or os.path.exists(diarize_abspath)):

does not make sense in general because we first check, if diarize_abspath was set or if the file exists. I think we cannot combine this with an or, because the second check will raise an error, if it's not set. This is what happened here, I guess. I think in thsi case we need a two step if clause, not an or.