ATrain Crashing

[thegian7]

I have been loving aTrain - until yesterday. For some reason startting yesterday, any time I try to use it to transcribe it crashes half way through. It doesn't provide an error.

[playtons]

Same here. Can't find any crashing events in Windows Event Viewer. It seems the problem was already mentioned in #9

[JuergenFleiss]

Same here. Can't find any crashing events in Windows Event Viewer. It seems the problem was already mentioned in #9

Currently we are unable to reproduce the error, also in #9

It's really strange for it to just start crashing. Did any of you reinstall it or change anything else? Does reinstalling help?

[playtons]

The problem seems to be fixed for now. The crash occurred only during the diarization step. When multi-speaker detection was disabled, aTrain functioned properly. Reinstalling aTrain alone didn't solve the problem. However, when I migrated to Conda to minimize dependency issues, I uninstalled Python, pip, and all wheel files, and started from scratch. This fixed the issue. Are aTrain's dependencies using the global Python environment or are they isolated?

[wenyuan-wu]

We have been using aTrain for quite a while, and it works almost every time.
We just used aTrain intensively to transcribe around 30 video recordings, and we have encountered many times crashing, and they appeared "randomly."
The video recordings are, on average, 40 minutes long in various formats like MP4, MOV, and MTS.
The hardware is a Windows 11 desktop with RTX 4070.
aTrain started to crash when transcribing the 2nd video. We have tried the following measures:

• convert video files into audio files (MP3 or WAV) and let aTrain transcribe the audio directly
• roll back nvidia graphic drives (NVIDIA Studio Driver - WHQL, from 555.99 back to 555.85)
• try different models (from medium to large-v2, or large-v1)
• with or without speaker detection
• delete "aTrain" folder in "Documents."
• reinstall aTrain via the installer downloaded from website, instead of MS store
• reboot system...
• random combination above

In the end, we managed to get all 30 videos transcribed successfully. However, the process was frustrating because we could not identify which measure above solved the problem, or not.
We want to share this experience here. We will further investigate if we can reproduce the problem and find out what could be the reason.

[JuergenFleiss]

Thank you for the testing already done and sorry for the frustration this caused you all. We were not able to produce crashes on various test systems, so it is a bit difficult for us to work on that.

One known issue is the GPU running out of memory for long files only, but this should not be an issue with a Desktop 4070 and its 12GB; our best testing card is a 2080 Ti with 11GB. Also with a 8GB card we never ran into this issue.

Generally, we are close to releasing the next version that will cleanly seperate the user interface from transcription pipeline. The pipeline is already available as a CLI here https://github.com/JuergenFleiss/atrain_core

What would greatly help, @wenyuan-wu , would be if you could test the transcription of your files with the new CLI. If the error is not reproducable there, it should be gone in the upcoming Desktop version.

[wenyuan-wu]

I have tried to reproduce such crashes, but it's difficult. The reason is that these video recordings are confidential and can only be examined on-site in a medical center, and it's hard to get an appointment...

My intuition (Bauchgefühl) is that such a problem is usually related to the file format and most likely occurs with video files. In many cases, the transcription was successful after the audio input was extracted from the video, allowing aTrain to work directly on the audio file.

I've tested the new CLI with our other interviews and they work fine. Although they can also be successfully transcribed using the current release (v1.1.0). There are two minor issues with the CLI which I have reported in the repo.

[Biga73]

At the first launch, almost 100% crushed
"delete "aTrain" folder in "Documents."" - Only this works. The program after that works fine.
The next day you have to repeat.

[Petertttt]

Chrashed here too (program closed towards the end of the process, without warning or saving) . Solved by deleting the Documents/ATrain folder every day once before use. Works stable now.

[JuergenFleiss]

Ok, thats difficult to debug actually. We will probably add the workaround to the FAQs for the time being.

Also, v1.2 has a completely rewritten backend, so maybe that took care of this issue.

[MikeHsu33]

I am not good at English, so this is translated.
My program kept crashing during execution.
I found online that adding an extra parameter,
temperature=0, to the transcription_model.transcribe function
inside def _perform_whisper_transcription solved the issue.
However, I don't understand why this works.
I am using aTrain_core.

transcription_segments, info = transcription_model.transcribe(
audio=audio_array,
vad_filter=True,
beam_size=5,
word_timestamps=True,
language=language,
max_new_tokens=max_new_tokens,
no_speech_threshold=0.6,
condition_on_previous_text=False,
temperature=0,
)