Checking AI Transcripts – Here’s What I’ve Learnt

Over the past few months we’ve been proofreading more AI transcripts than we have in the past.  Most of these have been good quality recordings, clear voices and accents, with no background noise, so I can see the sense in using AI to transcribe these files.  However, there’s still the need for human intervention, regardless of how good the interpretation from voice to text is.  Here’s some of the things I’ve noticed:

Formatting

Formatting is definitely the first thing that needs to be tackled if you want a useable transcript.  Most AI transcription software, particularly the auto-transcript within Teams and Zoom don’t give a readable format.  When the speaker pauses for breath, or to think about what they’re saying, the software assumes it’s a new speaker – so puts in a line break, timestamp and identifier, thereby breaking up the speech and making it difficult to read.  The first 15 minutes of proofing is therefore spent just sorting out the formatting so that the text looks coherent on the page.  Another bugbear is when the speakers change but it attributes half of the previous speaker to the new speaker, and vice versa. 

Grammar, punctuation and spelling

Following on from formatting, this then leads to issues with grammar and punctuation. With each erroneous break in the text, there’s a whole lot of punctuation that is either not required or is missing completely.  With each pause that triggers a new line break, it also inserts unnecessary stops, requiring the time intense removal of full stops and capital letters to join the sentence back together. 

Also, when we speak naturally, we don’t always take the time to pause our speech in line with written punctuation.  The AI transcript can therefore have huge swathes of paragraphs without any punctuation, which again leads to difficulty in reading and understanding the context and tone of how something has been said.  Human transcribers will pick up on these nuances and punctuate appropriately where required.

AI is definitely inconsistent when it comes to spelling – even the names of the participants in the interview can have umpteen different spelling versions throughout the document.  It doesn’t seem to use context to “cheque” that it’s using the right spelling either.  Again, these are all things that a human transcriptionist would address first time around.  If they don’t know the correct spelling they will highlight it for verification and be consistent throughout the document so that even if it’s not correct, a quick find and replace will resolve the issue.

Hallucinations

The hot topic when discussing AI technology is hallucinations, and I can see why.  The thing about AI transcription software is that it fools you in to believing it’s accurate.  As my dad used to say, if you say something with enough conviction, everyone will believe you, and AI does that brilliantly.  It cannot be trusted to have interpreted the text correctly.  It needs to be double checked against the original recording.  It will mishear words, it will be inconsistent with spelling, it will transcribe things that aren’t even said, it will attribute speech to the wrong person and, unlike a human transcriber, it won’t highlight any doubts or misgivings it has – because it doesn’t doubt itself, it believes it is always right.  I’m always amazed when I open an AI transcript and there is no inaudible text, and that worries me.  Having been doing audio transcription for 40+ years, there’s always inaudibles.  People don’t speak clearly; we use weak forms in our speech, we use connected speech that varies from language to language, we mumble, we turn away from the recorder, we hesitate and stumble over what we want to say, we use colloquialisms and dialect that AI has difficultly in deciphering.  And human transcribers will indicate this as they transcribe, marking things as inaudible or “sounds like” so it can be verified.  If AI can’t make it out, it will make it up.  My favourite is when it puts in “I love you”, just to soften us up. 

So, the big question we’re being asked at the moment is, is AI transcription quicker than human transcription – and thereby cheaper.  And while it may seem like that in some cases, there’s still an awful lot of human work needing done to ensure that an AI transcript is accurate and fit for purpose.  We’re not seeing a huge saving in time, and very often we’d wish we were transcribing from scratch as it feels so much quicker.  But we do recognise that AI is here to stay, and will improve with time, therefore we’re always available to proof AI transcripts when you don’t have the time to do so.

If you want us to help you with your AI transcripts, get in touch here.

Photo by Chris Montgomery on Unsplash