• After 15+ years, we've made a big change: Android Forums is now Early Bird Club. Learn more here.

Gboard's Voice-to-text bizarre inaccuracy

Kbarb

Lurker
Sep 4, 2020
3
0
Has anyone noticed how the whole Voice-to-text engine/algorithm seems to be oriented towards kids, teenagers, people that watch a lot of tv or watch & read entertainment news ?
As follows . . .

While you're typing, any word that seems remotely like a teenage band, pop idol, entertainment phenomenon, trend or trendy clothing, whatever - gets capitalized as such.
I'm talking about regular words that could be proper nouns (for some teenage reason), but are usually/often not.
Typically, for virtually every text I have to correct at least one or more such erroneous capitalizations.

Moreover, there are a lot of other random capitalizations, which don't make any sense at all.

Finally,
There are a lot of word guesses and insertions, which are often not even real words, when the obvious is not chosen.

It'd really be great if there was some control, choice, or personalization for this - like choices of algorithm versions / the adult vs teenage algorithms / insert only real words / Don't capitalize in the middle of a sentence / etc.

Anyone know how to leave feedback for the developers ?

I couldn't turn up anything except via the phone's settings for Gboard feedback, which is not really conducive to writing much.

Thanks !

Kent
 
Last edited:
(I'm not sure I have the right forum here actually - this about Google's Voice-to-text feature which I believe is part of GBoard keyboard. I'm actually using Microsoft SwiftKey, which indeed uses GBoard's Voice-to-text engine for that feature).

Anyway . . .

Has anyone noticed how the whole Voice-to-text engine/algorithm seems to be oriented towards kids, teenagers, people that watch a lot of tv or watch & read entertainment news ?
As follows . . .

While you're typing, anything that seems remotely like a teenage band, pop idol, entertainment phenomenon, trend or trendy clothing, whatever - gets capitalized as such.
Typically, for virtually every text I have to correct at least one or more such erroneous capitalizations.

Moreover, there are a lot of other random capitalizations, which don't make any sense at all.

Do you mean things like band names, clothing brands, etc. e.g. Nike sneakers, Gap hoodie, words that are proper nouns, and should be capitalised? If not, could you provide an example of exactly what you mean.

Finally,
There are a lot of word guesses and insertions, which are often not even real words, when the obvious is not chosen.

It'd really be great if there was some control, choice, or personalization for this - like choices of algorithm versions / the adult vs teenage algorithms / insert only real words / Don't capitalize in the middle of a sentence / etc.

Anyone know how to leave feedback for the developers ?

I couldn't turn up anything except via the phone's settings for Gboard feedback, which is not really conducive to writing much.

Thanks !

Kent

As I never use voice-to-text myself, I can't help you with that.
 
Upvote 0
So now of course I can't reproduce it, as it's an intermittent and unpredictable sort of thing.
I do remember I used the word "prince" once, and of course it had to capitalize that, even though it didn't make any sense in the context.

That's a perfect example actually - of words that could be proper nouns, but are usually/often not - but happen to be part of some pop band's name, whatever, and the algorithm decides to go with that.

Besides that there's a lot of random capitalization, and just absurd predictive word guesses being filled in.
Guesses that are not actual words, but sort of look like them.

I wish I knew who to report to as Google has very good developers, but it's surprising to me how bad this particular algorithm can be.
 
Last edited:
Upvote 0
In the past few months I have seen several posts and comments on VtT, mostly about punctuation more than capitalization but they have definitely been playing with something recently. Even so, and this I agree with, I keep hearing that Google's VtT is better than just about anyone else's offering.

As for your question about VtT and Gboard looking at the Play Store it would appear that they are one in the same, so sending feedback via Gboard should also send feedback about VtT. If you do not like the interface found through the app there is a feedback link on the app page at the Play Store under Developer contact with several ways to leave feedback.

On a side note I swipe and find that I occasionally get capitals for words that shouldn't be capitalized and incorrect words for what I feel should be the more common word. I rarely get incorrect words when I opt for VtT but when I do I wonder if it isn't my pronunciation more than an incorrect guess but often it's context as well. A lot of times I will have to put the word in context to get the correct word. For example if I say "whey" I get "way" so I'll have to say "curds and whey" then delete "curds and" to keep the "whey". I've also found that phonetically saying the word gets better results. For example not far from me there is a town everyone pronounces "bell-fountain" but to get the spelling correct I have to say "bell-fontaine" which is how it's spelled but not how it is pronounced. I get the same results when something is read to me for example, If Google reads to me "The ducks live in the pond." it will often/always say "live" like "is it live or is it Memorex" It doesn't read in context but seems to predict better when used in context.
 
Upvote 0
Interesting.

And so it's referred to as VtT I see.

I did leave feedback via my phone's app settings. Basically wrote it out first on an email and sent it to myself for copy & paste.
For leaving feedback via the Play Store, it looks like it's only the email contact apps-help@google.com, which doesn't look to be too specific.

The funny thing about context as you speak of, it doesn't appear to me that context makes too much difference. Perhaps some, at least you'd think anyway.
But a lot of times I'm thinking, "What made "you" think that word had any relation at all to the sentence it was in."

Another funny thing - sometimes it provides the wrong word, and I'll eff with it for while, trying all sorts of pronunciations to see if it can learn and get it right, just to see what happens. You'd think after the 5th or 6th time of getting rejected it would learn and try something new. But pretty much every time it just keeps providing the same mistake, (or occasionally different mistakes) over-and-over again.

That makes me think it's a somewhat static algorithm, not learning or AI, not very much anyway.

Btw, I think the processing is done locally - does anyone know ?
I've read Google's AI unit has been working on streamlined server based processing, but I don't think it's online yet, at least not for Gboard's VtT.
 
Last edited:
Upvote 0
Maybe there's something up with the whole system cause, while I'm not the greatest speller, I'll trace the word and sometimes even pause on each letter and still not get the word I need. That's usually when I resort to VtT. Can you just put in this one word I need, ugh.

Perhaps in a way it is as you described, it's taken on a younger vibe, and our vocabulary is so far removed it's feeding us words that get 2 million hits a day but not the ones that only get tens of thousands. I know that's reaching but if you're going to reach, reach big. Lol
 
Upvote 0
Do the kids even know about Prince?

If predictions are not personalised then of course they will be dominated by some average of the users of the system. I don't know whether this is the case as I don't use VtT.
I've also found that phonetically saying the word gets better results. For example not far from me there is a town everyone pronounces "bell-fountain" but to get the spelling correct I have to say "bell-fontaine" which is how it's spelled but not how it is pronounced.
Well to be fair it's not that likely that they'll know a local, non-phonetic pronounciation of a small town (I'm finding it puzzling that it's "Bell Fontaine" rather than "Belle Fontaine" too - the latter makes more sense and is linguistically consistent).

Anyway, be thankful you are from the same continent as the people who programmed the system: the problems become a lot worse if you aren't. ;) Though as I've had to say "AD" to get a guy in California to understand the word "eighty" I don't find it surprising that computers have trouble with accents.
 
Upvote 0

BEST TECH IN 2023

We've been tracking upcoming products and ranking the best tech since 2007. Thanks for trusting our opinion: we get rewarded through affiliate links that earn us a commission and we invite you to learn more about us.

Smartphones