Files longer than 30 minutes are noticable slower transcribed

[The-unknown-Shadowman]

The program noscribe works best with files that are no longer than 30 minutes. To optimize processing, files should be split into 30-minute chunks, processed individually, and then merged back into a single file at the end.

If files exceed this duration - especially in cases where they are as long as two hours - performance degradation and noticeable slowdowns may occur.

[kaixxx]

What system are you using? Mac? Windows? Cuda or not?

[The-unknown-Shadowman]

Both CPU and CUDA (1080ti) performance improves significantly when files are segmented into 30-minute chunks within a Windows 11 environment.

[gernophil]

This might be dependent on system and or video memory. How much does yours have?

[The-unknown-Shadowman]

32GB RAM and 11GB VRAM (1080ti)

[kaixxx]

A while ago, I did a test with a ~4h audio recording (~40 pages transcript). I've got the impression that with long files, the on-screen logging is bogging noScribe down because the text in the main textbox is getting too long. My plan is to limit the log to a certain length by discarding older parts of it.

This might also explain the slowdown you observe. However, splitting the file into chunks would not help in this case - as long as you don't restart the app and the log is still fully visible on screen (meaning, you can scroll up to see the log of the previous chunk).
You could test this by transcribing two 30 minutes chunks directly one after the other (use the "Start" and "Stop" parameters for that), then transcribe the second chunk again after restarting noScribe. The second transcription should be faster. (I'm not sure if 30 minutes chunks are enough to get a clear result because other processes on your machine might interfere, but it would be worth a try.)

[The-unknown-Shadowman]

At the moment I cut the file into 30 minute chunks and put them piece by piece into noscribe. I export them as txt and join them in cmd with type command.

[kaixxx]

put them piece by piece into noscribe

Do you restart noScribe in-between?

[The-unknown-Shadowman]

No, I don't restart between files. I just add them to noscribe one after another. However, as the process continues, the GUI becomes more unresponsive and the speed decreases. But the memory usage does not increase either in RAM or VRAM.

[kaixxx]

First, check your version that is shown in the main window right after startup. I suggest you use version 0.6.1, which includes some changes I made a while ago that will speed up noScribe with cuda.

No, I don't restart between files

This means that the growing log in the main text box (my suspicion above) is probably not the issue here. I'm running out of ideas at the moment...

While automatic chunking of the audio seems to be an obvious solution, it's not as easy as it seems:

• If you cut the audio in the middle of a word, you will lose part of the text at the transition point between chunks
• Also, the transcription of the second chunks misses the beginning of the sentence, which will also lead to a reduced quality at this point.
• Finally, whisper is internally using 30 seconds chunks anyway (with a sophisticated algorithm based on a sliding window with some overlap and the previous text as context).

[The-unknown-Shadowman]

No problem and thanks for your time. I coded a similar app myself. It cuts the files into 15 minute parts with a slight oberlap. Transcribes the parts and joins them at the end into one txt file. I use https://huggingface.co/deepdml/faster-whisper-large-v3-turbo-ct2 for transcription. It takes 15 minutes on a 1080ti to transcribe 125 minutes of recording.

[gernophil]

• Finally, whisper is internally using 30 seconds chunks anyway (with a sophisticated algorithm based on a sliding window with some overlap and the previous text as context).

Maybe there's a way to flush the memory after every of those steps then?

[The-unknown-Shadowman]

There seems to be a difference between internally processing 30 second chunks an loading the file for transcription. I think it's building up something like a context window over time or something similar. I really don't know in detail what's happening inside the framework or the model.

[kaixxx]

If you observe the same effect in your own script, it seems related to internal processing in faster-whisper. I would suggest talking to the folks over there, maybe they have a solution.