Why are Siri's speech-recognition algorithms so weird?
January 23, 2020 11:21 AM Subscribe
When I talk to Siri on my iPhone, her responses often make me wonder what the hell she is thinking. She comes up with some really weird interpretations of what I say. I'm not so much interested in fixing the problem (although that would certainly be nice) as I am in figuring out the reasons behind the idiosyncratic behavior.
I've had my iPhone for about two years, and I've noticed a distinct degradation in the quality of Siri's speech recognition. I prefer to speak my text messages, rather than type them in. (Incidentally, I'm a native speaker of American English, and I don't think I have any speech impediments). Here are some examples of what Siri has been doing:
For a period of a few weeks (or maybe longer), Siri would interpret normal, English words as if they were in a foreign language -- possibly German. I never bothered to do any Google searches on them, but I'd guess that the words were the names of towns or areas in Germany. The problem resolved itself.
I recently spoke a text message that said, "Pick up some ginger ale". The result was, "Pick up some ginger Al". Given that Apple's algorithms are supposed to pick the most-likely interpretation of ambiguous speech, I find this result to be especially surprising. Surely "ginger ale" is a much more common phrase than "ginger Al"?
Today, Siri spelled the name "Michelle" in two different ways within the same text: "She's going to find out from Michelle if Michele can make it tomorrow".
Siri has recently started changing past participles into present-tense verbs, with the word "it" following them: e.g., "I walk it home last night", instead of "I walked home last night".
Also lately, the word "message" has been coming out as "iMessage".
Siri sometimes types out a whole text correctly, only to automatically back-space over it, start it from scratch, and then interpret it incorrectly.
Anybody know why this might be happening?
I've had my iPhone for about two years, and I've noticed a distinct degradation in the quality of Siri's speech recognition. I prefer to speak my text messages, rather than type them in. (Incidentally, I'm a native speaker of American English, and I don't think I have any speech impediments). Here are some examples of what Siri has been doing:
For a period of a few weeks (or maybe longer), Siri would interpret normal, English words as if they were in a foreign language -- possibly German. I never bothered to do any Google searches on them, but I'd guess that the words were the names of towns or areas in Germany. The problem resolved itself.
I recently spoke a text message that said, "Pick up some ginger ale". The result was, "Pick up some ginger Al". Given that Apple's algorithms are supposed to pick the most-likely interpretation of ambiguous speech, I find this result to be especially surprising. Surely "ginger ale" is a much more common phrase than "ginger Al"?
Today, Siri spelled the name "Michelle" in two different ways within the same text: "She's going to find out from Michelle if Michele can make it tomorrow".
Siri has recently started changing past participles into present-tense verbs, with the word "it" following them: e.g., "I walk it home last night", instead of "I walked home last night".
Also lately, the word "message" has been coming out as "iMessage".
Siri sometimes types out a whole text correctly, only to automatically back-space over it, start it from scratch, and then interpret it incorrectly.
Anybody know why this might be happening?
Any chance your phone micrphone is blocked or clogged? That could degrade the quality of the speech recognition. You can find instructions online for cleaning out the iPhone microphone.
posted by mekily at 11:55 AM on January 23, 2020 [1 favorite]
posted by mekily at 11:55 AM on January 23, 2020 [1 favorite]
Some of these sound like it’s trying to identify common names in your communications. I can’t think of a good example offhand, but I’ve had some bizarre autocorrects to match speech to coworker’s names, sometimes breaking up a very common phrase as you saw with Ginger ale/Al.
Overall, I haven’t seen it degrade, though; most of its attempts to find proper nouns help me out.
posted by michaelh at 12:50 PM on January 23, 2020
Overall, I haven’t seen it degrade, though; most of its attempts to find proper nouns help me out.
posted by michaelh at 12:50 PM on January 23, 2020
This is just anecdata, but I spent a while trying out the built-in voice-to-text feature on my Mac last night, and it's astonishingly bad and weirdly unintuitive compared to my experience with Android. It autocorrects what I actually said, for example, into grammatically implausible full sentences, presumably based on likely adjacent words, without any regard for meaning over the course of an entire phrase. In my case, it seemed like it was trying to create full "units of thought," but missed or misrecognized every other word, and tripped up on moments when I rephrased myself halfway through a sentence. Even when it seemed to have gotten things mostly correct, just like you mentioned, it would go back and autocorrect the entire sentence into something absolutely incomprehensible. The end result was absolute gibberish, sentence after sentence.
I use voice-to-text on my Android phone regularly, and have a much different experience: it catches more or less everything I say, as I say it (imperfections and all), without trying to correct my grammar or autocorrect an entire phrase based on words that come at the end of it. The only thing it doesn't do perfectly are proper names it doesn't know and punctuation (speaking "period" will insert an actual period, whereas speaking "open quotation mark" just transcribes that compound noun-ish thing). I've started to use it to transcribe emails, since speaking doesn't wear me out half as much as typing, and I'd rather spend time at a keyboard writing something that demands craft choices spoken language isn't quite suited for.
posted by tapir-whorf at 1:47 PM on January 23, 2020 [1 favorite]
I use voice-to-text on my Android phone regularly, and have a much different experience: it catches more or less everything I say, as I say it (imperfections and all), without trying to correct my grammar or autocorrect an entire phrase based on words that come at the end of it. The only thing it doesn't do perfectly are proper names it doesn't know and punctuation (speaking "period" will insert an actual period, whereas speaking "open quotation mark" just transcribes that compound noun-ish thing). I've started to use it to transcribe emails, since speaking doesn't wear me out half as much as typing, and I'd rather spend time at a keyboard writing something that demands craft choices spoken language isn't quite suited for.
posted by tapir-whorf at 1:47 PM on January 23, 2020 [1 favorite]
Have you tried turning off autocorrect? I have noticed when typing that lately the iPhone autocorrect has seemingly gotten worse and been doing a lot of the kinds of things you describe. That is, maybe the issue isn't in Siri not recognizing what you are saying, but overzealous correction by the autocorrect algorithm after the fact.
posted by misskaz at 2:06 PM on January 23, 2020 [4 favorites]
posted by misskaz at 2:06 PM on January 23, 2020 [4 favorites]
This thread is closed to new comments.
posted by acidnova at 11:27 AM on January 23, 2020