Friday, May 13, 2022

Review of Gboard

 First of all, I have to say that this project has to do with getting some work done while I'm out on a walk. Therefore I need something that doesn't require a data connection.  With the price of data plans in Canada, I need to restrict my use of data.  One of my phones does have a data plan, but it's quite minimal, and the other has no data plan at all.  So I can only do data when in a Wi-Fi area or hotspot, which is generally not when I'm out walking.

In any case, this is a review of Gboard from Google.  But, as usual, I'm hacking the system, and trying to make it do things it was probably never intended for.  (I've been recommended to use the Otter program by a friend, but that program, I am quite sure, does require the use of data.  I hope to get around to testing it soon.)  However, the girls recommended that I do dictation on my phone.  I didn't think I had dictation capability on my phone.  But I found that one of the phones had Gboard installed on it. (I was rather surprised at this because most of the time the phones are fairly identical.)  I tried Gboard, and found that there was a control for switching on dictation, and lo and behold I could do dictation!

My first test was in an area that had wifi.  However as I tried with the other phone (that didn't have a data plan) and walked away from the house, I did find that the dictation suddenly stopped.  And so Gboard does require a data connection and cannot be used completely on its own.  Like Siri, Alexa, and all the others of that ilk, it has to pass the verbal stream to a back end processor for extraction of the text.

I did try a test using wifi hotspots on one of my walks.  This worked out well in an urban area, where there are a lot of hot spots, but probably doesn't work quite so well walking over the Alex Fraser Bridge or along the greenway trails down to Mud Bay.

It is reasonably accurate in many cases, but it does get easily confused.  Sometimes it is easy to see how it decided on something.  If you don't know the specifics of the local "Deltassist" office, it is easy to understand why Gboard thought it might be "Delta cyst."  (Then again, it obviously figured out that I was in Delta, since it capitalized it.  So why doesn't it know about the office?  Although in another case is decided on "Delt assist."  Does Google think I'm a frat boy?)  I never did figure out why it decided on "recluse" instead of "reclose."  (Interestingly, Google lists Gboard as Gboard, but GBoard transcribes Gboard as "gboard.")  But in cases where you repeat material, say three or four times to make emphasis, Gboard will eliminate the multiple references.  In other cases it just simply gets confused and refuses to take down any dictation and text of anything.  Sometimes Gboard will, reasonably accurately, take down a phrase, and then suddenly decide that you didn't mean that, and delete the material it has just transcribed.  Most annoying of all is when it does this for three or four whole sentences, or even an entire paragraph, and then delete the whole thing, and you have to start again from scratch!

On this initial test I was primarily using Gboard to dictate messages into WhatsApp.  I was mostly doing quick texts into WhatsApp, or, at least, not very lengthy dictations.  Interestingly, on the Samsung Notes app, saying the command "new paragraph" generates a paragraph.  Sometimes.  (About half the time.)  Sometimes you just get the text "new paragraph" in the text.  (Pretty much always, in WhatsApp.)

Gboard doesn't seem to be able to handle more than about a minute at a time. Therefore short dictations into apps does seem to be its forte.

I did want to try dictating a longer piece, but didn't necessarily want to take time out from my walk to find a hotspot and sit and dictate for a while.  What I did was use the voice recorder to record, in several chunks, a piece of about ten minutes long.  This was an article on the choice of virtual platform for conferences.  The article dictation I then played from one phone to the other. This allowed me to test how long Google was actually able to listen and record dictation.

The effective length of any dictation did seem to last only about a minute.  I do not know if this is an actual limitation on the program, or if it is simply an issue of buffering.  The dictation did stop in many places.

There do seem to be some phrases that Gboard has a problem with.  Gboard will convert the phrase "most of the time the phones are pretty identical", but then immediately deletes it.  Every time.  Also the phrase "one of the phones had gboard installed on it."  I don't know if I'm just getting paranoid, but Gboard seems to refuse to take down any text that is a reference to Google or its products in less than glowing terms.  With the voice recorded piece I would play and replay the section that was problematic, and Google would still not take down the text of the dictation.  Gboard has problems with any phrase containing the word "buffering" which is *EXTREMELY* interesting.  Sometimes Gboard simply stops transcribing your work.  In any case there can be large chunks of the text that Google misses.  (Gboard is definitely *NOT* recommended for the visually impaired: you *must* check what it takes down.)

Background noise *can* be an issue.  I dictated the piece about virtual platforms for conferences while I was walking by a major intersection on Nordel Way.  Even so, Gboard was able to transcribe that section fairly well.  In a later test, transcribing a voice recording at home while filling a tub, traffic noise was not problem, but water filling a bathtub was.  However, another test, voice recorded while walking along the approaches to the Alex Fraser Bridge, was impossible to transcribe, so, at some point, traffic noise can be a problem.

No comments:

Post a Comment