News for hackers. Stuff that matters.
Re:Automatically transcribe an interview, meeting or video by spzb on Monday May 10, 2021 @05:33AM (#27104239)
In my experience, they're pretty poor in English too. Especially when it's ad hoc conversation where people don't finish sentences, repeat themselves, "um" and "err" etc
Re:Automatically transcribe an interview, meeting or video by hnbad on Monday May 10, 2021 @06:13AM (#27104404)
A problem most people don't think about when talking about transcription is that people don't talk like books. Not only do you get unfinished sentences and filler words, you also get garbled words, non-standard pronunciation, and so on.

In the case of pronunciation this primarily poses a problem with detecting the intended word, but in other cases "cleaning up" the output may lose contextual information (e.g. what a speaker was going to say before cutting themselves off and using a different word). This is difficult enough for a human to get right, let alone a machine.

Re:Automatically transcribe an interview, meeting or video by rob74 on Monday May 10, 2021 @08:26AM (#27105214)
Also, when transcription programs (most familiar example: YouTube) fail, they usually fail on the words that a human listener would also have trouble understanding / telling apart. So the transcription is useful if you are deaf or forced to watch the video without sound, but if you're using subtitles because your English is not good enough to understand the speakers without them, their usefulness is pretty limited...
Re:Automatically transcribe an interview, meeting or video by hobofan on Monday May 10, 2021 @08:32AM (#27105265)
Once you go beyond ~7 words (= what people would utter to their virtual assistant), the quality of all off the shelf tools (both open source and offered services) is laughable. Sentence boundary detection, punctuation, speaker segmentation, and all those features you would need for good transcription are in a really bad state.
Re:Automatically transcribe an interview, meeting or video by lostinthefield on Monday May 10, 2021 @11:53AM (#27107422)
"Really bad" is an exaggeration, I think. The auto-transcription features in both Google Meet and Zoom are more than acceptable, they're often very useful in catching missed words during a meeting.

They trip up on technical jargon but handle everyday conversations just fine, including speaker detection, punctuation, idioms, etc.

But that's also a slightly different use case, where each speaker is in their own (somewhat) quiet environment and on separate connections (and thus audio tracks).

It's much harder to do all that after the fact, like with a recorded video.

I find Trint.com, which is partially automatic, to be good for that... the AI does a first pass, and a human cleans it up afterward. YouTube has a similar assisted-auto feature for their captions, minus speaker separation.