Auto transcription is completely different to static transcription

[Dominic Witherow]

Issue:

The auto-transcript of my sequence is gibberish (ie just random letters with no actual words in my chosen language - English), but the static transcription of the exact same sequence is almost 100% accurate. How is that even possible??? It must be a mad bug to make such an enormous difference.

Tech details:

PP 25.1.0 (build 73)

Mac OS 15.3

2024 MBP M4 Max 16" 128GB 4TB SSD

Sequence is 2592 x 1080p anamorphic

Footage is 5952 x 3968 anamorphic 10 bit HEVC 200mb/s LGOP shot on Panasonic S1H

48K 16bit Audio recorded on camera via external mics through Panasonic XLR adapter

Workflow: 

With the sequence on the timeline I click 'transcribe' and after a few minutes the transcription appears and it is mostly random letters, formatted as though following alongwith the speakers, but meaningless. The first line of the transcript reads: "Er ikke ham, for man bør. Riksmålet Almaglid er ikke skål og stooore skoskap og."

When I click on 'generate static transcript', the same line reads: "I come from a very small town. I feel like it's close to us. Close up. When Tom", which is about 90% accurate. Why would the same transcription be so different? In the video, the speaker is an elderly lady, whose first language is Spanish, but her English is perfectly clear and undertsandable, as demonstrated by the static transcript. Whereas the auto-transcript seems to think she is Danish (I'm guessing).

Reproduction:

As simple as described above - 

1. press transcribe

2. wait for transcription, which is gobbledegook

3. press 'static transcription'

4. await result, which seems pretty darn good

What should happen is I should get the same text for both transcriptions (ideally the accurate version for both!)

What does happen is I get meaningless crap from one and good from the other, but at the cost of the text-based editing functionality.

Thanks for your help!

[Rach McIntire]

Hi @Dominic Witherow,

Thanks for submitting your bug report.  

Do you have the language you want to set to default? Are you transcribing from the timeline's source clip? Check this thread to see if it is similar to this issue.

I hope we can help you soon. Sorry for the frustration! 

[Stan Jones]

@Dominic Witherow,

My questions are the same as @Rach McIntire's:

What is your default language? Go to Preferences -> Transcription -> default language option. Do you  have autodetect set by any chance?

When you click "transcribe" in response to putting an untranscribed clip on the timeline, it uses the "automatic" settings and does not give you the chance to pick a language. When you manually transcribe (e.g. from the source monitor/3 dots menu or right-clicking the Project Panel) and when you create a Static Transcription of the sequence, it asks you for the language.

There may be a problem, but it is not common, and the first thing to check is the default language settings.

Stan

[Dominic Witherow]

Thank you, both, so much! There was a mis-match in the language setting (Norwegian!!!) when using the source monitor. Set back to English and it all worked.

However, surely this is a feature error? There should be consistency between the various transcriptions? Also, how on Earth did Norwegian become the default language, when everything else is set to a different language? THAT stinks of a bug somewhere.

Anyway, it's working now, so I'm most grateful for your help. 

[Stan Jones]

@Dominic Witherow,

> There should be consistency between the various transcriptions? 

In my experience, this works correctly. The confusion is whether you are using a transcription workflow that uses the default preference with no opportunity to change it (automatic set to all imports or to clips added to sequences, or automatic when presented with the "transcribe" untranscribed) or a workflow where you get the language selection in the transcription options dialogue. The latter may be "sticky" and leave you thinking your preference is okay, even if it is not.

I think there are now 3 reports of similar behavior, but no method for replicating this as a bug. That doesn't mean there's not one....

Stan