Hello, I know we keep talking about transcripts and tagging speakers. But this is the first I’ve noticed where the transcript got it right, but when I regenerate the summary, incorrect speaker is credited with having said something.
Maybe because I didn’t identify all speakers?
See attached. Mark Wahlberg did not have anything to do with the letter to the Ephesians. That was someone else (I didn’t know their name so I didn’t tag them).
FYI this is from the “Hallow” app in case anyone was confused.
I am seeing the same thing. The speakers are tagged appropriately in transcript but summary incorectly associates seemingly all the time. It is frustrating because it makes summary less valuable. I saw on several models and different meetings. Maybe I am just doing it wrong.
Hey guys, this normally happens when a specific segment might have been mistagged by the AI. We are working on improving this to make sure it has the best performance possible but in such a case we suggest using our transcript edit feature to tag the proper speaker and then regenerate so it gives the best possible output.
This happened to me today and I found a bug in the Android app. When you “add speaker” and select someone and come back out to the transcript, the speaker is briefly applied to the segment in question, however it is then immediately moved to previous segment thus tagging incorrectly.
How can we product a correct summary with this issue in place?
Great, glad to hear! So how do I resolve the switched speakers in the summary now? I’ve tried regenerating the summary a few times and it’s still always switched, while the transcript appears correctly.
I am having a similar issue, but my transcripts are interpreting the information, speakers incorrectly. For example, if I have a call with someone, it transcribes the other party as me. Is this a setting issue?
Thanks for reporting this! If the transcript has the speakers mixed up (transcribing the other party as you), this is a speaker diarization issue from the AI model. Unfortunately there isn’t a setting to fix this - we’re actively researching and improving our models in the backend to reduce these errors. For now, the edit transcript feature is the best way to correct these mistakes when they happen.
I’ve had similar problems. However, I’ve also tested one other recorder and three other phone apps. They all have the same problem. Notion’s AI, which I’ve been using for transcribing meetings for months, does it, too. I’m beginning to think it may be a limitation of the current technology.
