Dictation hallucinations

This has happened several times to me in the past couple of months.

Today I’m dictating a text message. It’s about three short-ish sentences long, and I’m watching it as it goes. Everything is looking good. Then I stop the dictation and click to send. Apparently in the process of stopping and sending it retconned my message by deleting half a dozen contiguous words, leaving (a) a text message without crucial information, and (b) a sentence that’s semantic nonsense.

There are TONS of articles about how to disable dictation, how to disable auto-capitalization, how to disable predictive text, etc. None of that seems to be what I want.

I want Siri to, when I stop dictation, not screw with the message - especially not something more than a full sentence ago. It seems like they’re doing some sort of AI processing at the end of the whole message, and the fact that I have to re-read everything I dictate multiple times is incredibly annoying.

Does anybody else have this problem? Is there a fix?

Dictation may have misheard what you said on a word-by-word basis but it tries to correct itself based on the context provided by words that follow. That’s how it works. Not to say that it won’t screw up sometimes based on how you pronounce words, pause, or try to make a correction. Don’t hit “click to send.” Turn off dictation by clicking the microphone icon or using F5 on your Mac’s keyboard.

Yes, it happens almost every time I use Siri dictation. So I only use it when texting family.

I see that, but…a dozen words later? And deleting six-word phrases? Does it think that I said nothing for a dozen syllables of dictation?

@webwalrus, you’re a programmer. I would expect you to have a stronger understanding of the difficulties inherent in the process of getting user input by “listening” and then correctly interpreting what was said. It is possible that someone observing you could offer suggestions on why your dictation results are less than what you expect. Intonation, speaking speed, accent, pauses, mumbling, and attempted corrections, for example, all contribute to mis-hearing by a machine/software interface. My advice is to gather your thoughts before speaking and to be more self-aware as you speak. And adjust your expectations to reduce your apparent frustration. Apple’s marketing wizards encourage us to expect near perfection. Don’t fall for it.

Well that’s the thing - I am a programmer. When I say a sentence and it gets it wrong, I get it. Usually it’s pretty close, or I can see what it thought I was saying/trying to say. And I can fix it. This is why I watch it as I’m dictating, and edit/adjust/stop dictation/etc. as needed to get things right. I understand that dictation is potentially fraught with errors.

But…

Looking back on my estimate, I was a little off. This text was about 40 words. I watched every sentence as it was dictated. It was absolutely, 100%, perfectly correct when I successively pressed the button to stop dictation and then pressed the button to send.

But sometime between the “stop” and the “send,” it suddenly decided that half a dozen of the first ten words - which it had already heard and transcribed properly, and I had mentally verified as correct - suddenly didn’t exist anymore. It didn’t re-word them. It didn’t reformat the sentence for clarity. It deleted six words, leaving a sentence that was utter gibberish without them.

As a programmer, it looks to me as if, when the “stop dictation” button is pressed, something is doing a final lookback at the complete results of the dictation and second-guessing itself - and doing a horrible job of it. I’m wondering if there’s a way to disable that behavior without either giving up on dictation entirely or having to re-proofread a text message I’ve already proofread. :slight_smile:

1 Like

Perhaps you are running into a length limitation (buffer size) if words at the beginning are being chopped?

I think that this is unfair. @webwalrus has said that Siri correctly dictated what he said and THEN went back and changed things.

It’s done this to me too and it makes no sense.

2 Likes

Programmers know that “things that make no sense” actually do turn out to have a reason behind them.

I’m with Geoff and Walrus here. The behavior is not as expected – wether you are a programmer or not.

I’d take this to Apple Feedback, and maybe to Apple Discussions.

Then file a bug report with Apple!

This forum is often used to vent one’s frustration with Apple and its hardware and software. And I think venting frustration is a useful behavior.

But a more practical response is to either adapt one’s behavior to what’s currently being offered by Apple or to report it and see if you can convince Apple to change (aka “fix”) their system.

1 Like

I could buy into the idea of a buffer issue, except it wasn’t the words all the way at the beginning. The first handful of words were fine, then it deleted half a dozen words. It would still be unacceptable, but it would be an explanation.

I’m sure there’s a reason. That said, (a) it doesn’t jump out at me, and (b) I don’t recall this particular issue happening prior to the current major version of iOS. Don’t get me wrong; I had lots of problems with dictation before, but my memory is that they were problems getting words onto the screen. Once the words were on the screen, it was fine.

I am therefore guessing this is some sort of new post-dictation (AI/ML/whatever Apple is calling it these days) feature that’s misfiring.

Or you can attempt to gather more information about the problem by determining if it’s happening to anybody else, and/or try to find a workaround for this (seemingly) new feature. Which is what I actually did. :slight_smile:

I’m really hoping for a toggle somewhere where this new behavior can be shut off. I’m guessing that I’m hoping in vain.

4 Likes

+1


This thread turned weirdly aggressive, let’s all take a breath! Malfunctioning tech (as in ‘a failure to work as expected’) is frustrating and we’re allowed (encouraged even!) to seek consolation from fellow humans even if the answer is simply “file a ticket”.

The new text input software does seem a bit odd in places.

I have watched Drafts Dictation go back and edit what I’ve said, but so far the corrections have been correct, and I’ve only seen it do it for the preceding sentence, not a sentence (or sentences) before. I speak to the dictation software like a BBC news reader though (does that joke work outside the UK? We’re about to find out!) and much slower than I talk in regular speech, as I’m new to dictation and don’t trust the software that much (plus having watched less sophisticated software struggle for years with regional and “English as a foreign language” accents, I’m wary).

I have watched Apple Messages incorrectly “correct” messages I’ve typed. In a message last week it corrected a correct word to one that wasn’t even a word after I was already further along in the sentence. Or if it was a word, it’s not a word in use any more. Baffling. Except I was so amused my friend and I used the incorrect word for the rest of the conversation and now my phone thinks it is a word. Which leads me to the point I was going to make: my thought with this glitch you’re experiencing is that your device is “learning” from the repeated deletion and thinks it’s expected behaviour?

I read somewhere that part of the “problem” with the big language update in iOS17 is that because it’s basically new (a new program rather than an upgrade), our devices are having to learn our weird quirks all over again. So if you always pronounce a word wrong, for example, or you always hit q instead of w on your iPhone keyboard, your phone knew that before and now has to learn it again. So it will improve over time once it relearns all those things.

2 Likes

That might be the best answer I’ve seen so far. It sees me routinely delete several words of its gibberish, can’t identify a pattern, and figures that I must just randomly want to back up a couple of sentences and delete things.

That would be the most hilarious thing in the world to inadvertently train an AI to do – indiscriminate, random deleting :slight_smile:

1 Like

I’m not young in body and I’m not old at heart. But, I’m still fairly fast of mind.

The hardest part of dictating (which I use a LOT on my iPhone 14) is relearning to stop myself before I send something, wait for a beat or two and look at what is typed.

Trips me up more than I care to admit.

3 Likes

I gave up trying to do this years ago and started composing messages/emails in a separate app.

1 Like

Is this in mimestream?

If so, it’s a known bug.
And it’s infuriating when it happens.

Exactly this. +1

Why else do we have a forum…?

1 Like