0

The Whisper ASR is a very accurate speech recognition system, on average. But in some rare cases, it has very bad output.

E.g.:

ground truth whisper output
seven damn it!
her jewelery shimmered Hey, did you lose your mind?
the englishman said nothing Is that really what he said? I don't know.
he lived in charles street surrounded by his workers houses He leave in ambulance before,.. I hope you enjoyed this video, and I will see you in the next video. I hope you enjoyed this video, and I will see you in the next video.
we won't talk about her any more if you'd rather not we indeed We won't talk about her anymore if you'd rather not. Be sure to let us know in the comments below what you think of this video, and if you have
otorhinolaryngology Oh, I know that I'm gonna die.
ices ISIS
air each and Air富兵 Air equation Air per unit Air air FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR FOR
tranced Transcribed by ESO. Translated by —
generous I think making love with differences is important for amore. This is trying to explore something that the situation is not clear, I think forgetting that too many differences다른 것들에red get divided too quickly Having the opposite Hey!
dec one, two, three, four, five, six, seven, eight, nine, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38,
car What? I'm here now. I just... I'm... I just got a call. Hi. Hi. Hi. Yeah, hi. Hi. Hi. Hi. I'm here now. Hi. Hi. Hi. Hi. Hi. Hi. Hi. Hi. Hi. Hi. Hi. Hi. Hi. Hi. Hi. Hi.
two 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 pass 1, pass 2, pass 3, pass 4, pass 5, pass 6,ador, pass 2, pass 3, pass 4, pass 5 index cross tachometer to Thinking eating onion
snake BeGitorne is my daughter. She loves cake. So before I turn it over to her, please, close with a circle. And next, taken from New York and cheers, Now is whose turn is it. Let's see from the back. Yeah,
other inside? Is that right? Can you hear it? No, okay, it's too loud. Awesome. effective type? Yes. Peripheral anxiety. Okay, what are we going to do? Yeah Exactly. I think something. So selfish. I don't know whether I want to
validity Can you say these words again for me? Why don't you go with your body? I, why don't you go with your mind? Your body is very different when it goes away, your Catholic body goes away. missed these words.
edit Did I forget something useful? Did I forget something useful? Did I forget something useful? Did I forget something useful?? Did I forget something useful? Did I forget something useful?? Did I forget something useful?? Did I forget something useful??
joseph Joseph So I hope you are all feverish about the axis of the세 punched bottle. It was written by the Italian Karl Ernst Strauss who said, The progress of our human朱 is beginning of the sigma informatic body to achieve a useful purpose
streak Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank Sc blank oblivious
couples Pupvals. Please turn the sound off for a minute and we'll continue the rest of the Zoom conference. If you can't hear me well could you direct your volume down to 5 zero and that way we can hear the chair and tradition.
translations But it's important to have a clear vision of the future. And basically, that's really all I'm going to say. Thanks. Thanks, everyone. Thank you for your time. Thanks. Thanks, everyone. Thanks, everyone. Thanks, everyone. Thanks, everyone. Thanks, everyone. Thanks, everyone. Thank you. Thanks, everyone.
partnerships That's the end of the video. Thank you for watching. Lots of heat, December. Keep writing your comments below for new videos and feel free to contribute. If you have any questions, feel free to ask away, or make comments, post any of your comments. I'll see you next week.
nine Now, I'm going to show you how to do a couple of things. So, I'm going to go ahead and create a new project. I'm going to create an image. I'm going to create an image, and I'm going to create a new project. So, I'm going to create a new project. And I'm going to create a new project.
quote 靠 advantages. Oh, 안녕하세요ES represent Sorry, W represent Sorry, W represent Oh, I represent Oh, I represent Oh, I represent Oh, I represent Oh, I represent I represent Oh, I represent Oh, I represent Oh, I represent Oh, I represent Oh, Oh, I represent Oh, I represent Oh, I represent What is I represent What is a I am I represent Oh,
speaking Wait one minute. I think it's working. Just a second. I can see when I just pressed it. All right, try it now, but we're only going to press it for about a short while. So if you press the home button down here now, it says pop guards. So like from four to four. So with that I can pull that up. So that there are two of those.
appointment Conjure Hrica Conjure Hrica Conjure Hrica Conjure overtek Conjure Hrica Conjure overtek Conjure Hrica Conjure overtek Conjure overtek multi Conjure Hrica Conjure overtek multi Conjure Hrica Conjure Hstä Conjure Hkam Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Conjure Hka Kah Conjure Hka
frankfurt Encrypt. Let website bootselect. This is a code now. Let's see what happens if I prevent it from working here. loves Sangio. To start the Partition manager, message the first character and then mark the fullchaap in a time sheet. Set the global filter time, go to britishfamilies right click and choose the default feature you will choose lets see if it works. no it's not. While we dynamics with time changes, the Bartod uses the same doctrine of repetitions to make Emergency society work. We will oktrish. Save it and we can exit ask if it doesn't work uninst если neat revolve on az. Next we should replace the date on the property tag with date if not we are więkshee and we will loop. Save the way. Next we Sena considered are the object and values that lit the traffic to the present

It clearly overrelies on the language model, but why does it rely on it way too excessively in some rare cases?

5
  • 1
    Looks like mistakes in its training data to me.
    – Draconis
    Oct 22, 2022 at 0:25
  • 1
    Are these replicable errors (digitized input)?
    – user6726
    Oct 22, 2022 at 0:27
  • @user6726 I believe Whisper's inference is fully deterministic. Oct 22, 2022 at 1:05
  • @Draconis training data is noisy indeed, but contains 680k hours of audio files. It's interesting that in some rare cases it switches to some "catastrophic" mode. Oct 22, 2022 at 1:06
  • @FranckDernoncourt, the question is not about the abstract theory of the program, it is about whether the putative acoustic data is fundamentally defective. Can anyone besides you verify these claims?
    – user6726
    Oct 22, 2022 at 4:23

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Browse other questions tagged or ask your own question.