Yet Another Cycling Forum

General Category => The Knowledge => Ctrl-Alt-Del => Topic started by: mrcharly-YHT on 14 November, 2021, 12:17:40 pm

Title: Speech to text - pixel 6
Post by: mrcharly-YHT on 14 November, 2021, 12:17:40 pm
I've used speech to text on a Sony phone. When I injured my arm. It was ok but lots of errors and slow.
I now have a pixel 6. reportedly the chip is built specifically to handle speech to text.

I have to say it's absolutely incredible. I've tried it when entirely disconnected from the internet and it still works perfectly. I'm now using it for some work when I want to make notes and not be distracted by keyboard or my terrible handwriting .

I would be really interested to hear barakta's  opinion of this in operation.

Possibly the downside is I'm now starting to write overly long responses to things on Facebook, messenger, WhatsApp, etc.

This is quite a game changer. The first generation of chips designed to handle speech to text and it is very good. If this continues chips get better, I can imagine that keyboard use will decline hugely. Computers will then probably shift to tablet format since a keyboard will be superfluous.
Title: Re: Speech to text - pixel 6
Post by: barakta on 14 November, 2021, 01:07:25 pm
It would be interesting to see if the Pixel 6 could cope with my speech.

Doing it in hardware is interesting. I wonder if the traditional "Dragon" based software packages are now getting more and more behind the times.

I haven't tried a new text-to-speech since pre March 2020 as that's when I was last at work and having the opportunity.
Title: Re: Speech to text - pixel 6
Post by: Polar Bear on 14 November, 2021, 01:24:17 pm
I have been considering upgrading the pocket slab for a while now.  The speech to text claims for the Pixel 6 and 6 Pro make them a really tempting proposition. 

They do the other things adequately well too and the Pro has a reasonable sized screen which would be an upgrade to my current daily driver, and, I'd be shot of samsung's bloatware.

Very tempting.  Very tempting indeed.
Title: Re: Speech to text - pixel 6
Post by: mrcharly-YHT on 14 November, 2021, 01:52:11 pm
There is a lot of software involved in the speech to text translation. The point about the hardware is that he has been optimised to cope with the software load . I haven't seen the full internal specs for the chip but knowing from the design end some of the architecture that went into it. I suspect there is a lot of machine learning aspects to this. Some of the new parts of the architecture are there to optimise  large quantities of mathematical operations. This is expensive in terms of silicon so there's no reason to add it unless you're using machine learning a lot.

I find that the conversion works best if I speak in a measured cadence. It has made me become aware of my diction and realise that I have a habit of speeding up when saying some combinations of words.
Title: Re: Speech to text - pixel 6
Post by: Cudzoziemiec on 14 November, 2021, 01:54:15 pm
I thought the move was to AI / machine learning so speech to text hardware is interesting. When you say it "works perfectly" what sort of text does it give you? Perfect for Facebook but is it perfect, after proofing, for something more professional? How does it cope with homophones? Background noise? Can it distinguish change of speaker? Other languages?
Title: Re: Speech to text - pixel 6
Post by: Ham on 14 November, 2021, 02:35:15 pm
Can't give you all the answers, but I can tell you that homophones are mostly dealt with very well, based on context. I can also say that is far superior to the speech to text in the normal Google assistant operation, which is handled remotely I think. You should have heard me talking to the little b*ast*rd last night:

"Ok Google, play me some Barney Kessel"

"Sure, playing Barney Castle....."

{Repeat ad nauseum}

Background noise is dealt with well, too
Title: Re: Speech to text - pixel 6
Post by: Mr Larrington on 14 November, 2021, 05:57:01 pm
Mr Larrington:Alexa! Play Some Barney Kessel!
Alexa:Booking your eye test...

The tatty black one with a stain up the front, please.
Title: Re: Speech to text - pixel 6
Post by: chrisbainbridge on 14 November, 2021, 07:52:38 pm
Interesting as naturally speaking requires significant processing power and RAM to give a good response.  I reckon I am about 98% correct.  Biggest bugbear is that it prefers period to full-stop and will frequently turn it to will stop.  This will happen about 5 times in 20 pages of double spaced text so not bad.

My biggest problem is that I cannot proof read my own work for about a week, then suddenly all the mistakes jump out at me!  My PA therefore reads all my reports and sends them back to me.  We now have it streamlined into a paperless office where it gets dictated, proofed, corrected and emailed.

Title: Re: Speech to text - pixel 6
Post by: Cudzoziemiec on 14 November, 2021, 08:16:19 pm
Quote from: chrisbainbridge on 14 November, 2021, 07:52:38 pm
Interesting as naturally speaking requires significant processing power and RAM to give a good response.  I reckon I am about 98% correct.  Biggest bugbear is that it prefers period to full-stop and will frequently turn it to will stop.  This will happen about 5 times in 20 pages of double spaced text so not bad.

My biggest problem is that I cannot proof read my own work for about a week, then suddenly all the mistakes jump out at me!  My PA therefore reads all my reports and sends them back to me.  We now have it streamlined into a paperless office where it gets dictated, proofed, corrected and emailed.
So you have to tell it what punctuation to put in, rather than it interpreting punctuation from your speech patterns. A bit like dictating a letter to a secretary circa 1974! So it wouldn't cope with a conversation.
Title: Re: Speech to text - pixel 6
Post by: Kim on 14 November, 2021, 09:12:07 pm
Dragon was always optimised for dictation / control of a computer by a single speaker (with a decent microphone, so no background noise), rather than more general transcription.  Until relatively recently it was the only thing that was good enough to be worthwhile, but as computing power has increased it seems to have been leapfrogged by the machine learning approach.
Title: Re: Speech to text - pixel 6
Post by: chrisbainbridge on 14 November, 2021, 10:18:37 pm
Quote from: Cudzoziemiec on 14 November, 2021, 08:16:19 pm
Quote from: chrisbainbridge on 14 November, 2021, 07:52:38 pm
Interesting as naturally speaking requires significant processing power and RAM to give a good response.  I reckon I am about 98% correct.  Biggest bugbear is that it prefers period to full-stop and will frequently turn it to will stop.  This will happen about 5 times in 20 pages of double spaced text so not bad.

My biggest problem is that I cannot proof read my own work for about a week, then suddenly all the mistakes jump out at me!  My PA therefore reads all my reports and sends them back to me.  We now have it streamlined into a paperless office where it gets dictated, proofed, corrected and emailed.
So you have to tell it what punctuation to put in, rather than it interpreting punctuation from your speech patterns. A bit like dictating a letter to a secretary circa 1974! So it wouldn't cope with a conversation.
I think you can get it to auto punctuate but always preferred to do my own. I have been using it for about 8-9 years so pretty happy with it. I have about 20-30 commands and macros to open files, save, send, print. Auto text is easy, anything from a signature to several pages. Search and replace through a document.

I produce 4-5, 20-25 page reports per week with associated letters.
I totally replace the computer every 3 years and restart the audio learning as uncorrected mistakes eventually clog the system.
It has now been bought out and it here has not been an update for a number of years sadly.
Title: Re: Speech to text - pixel 6
Post by: mrcharly-YHT on 14 November, 2021, 10:30:15 pm
Giving this a go with background noise so music and speech from a TV program. It seems to be coping very well. I'm not using a microphone, just holding the phone in front of  me.
Title: Re: Speech to text - pixel 6
Post by: barakta on 15 November, 2021, 12:04:50 am
I specifically am interested in how the Pixel 6 will handle my speech as Dragon hates it. My DSA assessor colleagues all use the same profile that one of them created when demoing Dragon to students and it works. Despite training extensively, Dragon will not pick up my speech reliably and trying to speak at a computer makes my speech deteriorate.

I am in awe of people who can make Dragon work for lots of writing, especially as my medical stuff that is dictated tends to be rambly, repetitive and not well structured. I guess if it was a standard report format it might work better. There is a definite art to dictation rather than writing or typing.
Title: Re: Speech to text - pixel 6
Post by: Ham on 15 November, 2021, 08:14:10 am
Happy to set up a virtual session to experiment, if you fancy.  I've no idea if the VoIP will degrade the process but it might be worth a go