Conversation

Notices

1. halcy​:icosahedron: (halcy@icosahedron.website)'s status on Friday, 18-Jan-2019 03:54:55 JST halcy​:icosahedron:

   • Woof?

   @Doggo one thing that you can do (if you haven't already) is to turn on and train (or turn of and then back on) the "hey siri" feature. I'm not 100% on if apple uses the voice print from that to improve their general speech recognizers performance, but that's a thing that they could do (and they probably do it).

   Beyond that, it's probably very hard. The in the cloud nature of how siri / g voice / alexa work makes adaptation to specific users hard (usually what is done instead is that a

   • halcy​:icosahedron: (halcy@icosahedron.website)'s status on Friday, 18-Jan-2019 03:57:03 JST halcy​:icosahedron:

     • Woof?

     @Doggo type of abstraction of how you speak is passed into the recognition system as an additional input so that the system can work well for different kinds of voices)

     same goes for specific words - though you CAN teach siri how to pronounce them, that's likely mostly for output

     generally, since it's a machine learning system, what you want is for your voice to be as close to what the system was trained with, which is why using a microphone might not actually be helpful, but if you

   • halcy​:icosahedron: (halcy@icosahedron.website)'s status on Friday, 18-Jan-2019 03:59:00 JST halcy​:icosahedron:

     in reply to

     • Woof?

     @Doggo notice that a certain distance from the microphone works or doesn't work very well, that might not just be your imagination.

     For the Far Future, you could try to contribute recordings to https://voice.mozilla.org/en - that's not going to help you immediately, but if people or companies start using mozillas database as part of training their systems, your voice would be in there - and a system having seen your voice specifically is, of course, ideal

   • halcy​:icosahedron: (halcy@icosahedron.website)'s status on Friday, 18-Jan-2019 03:59:52 JST halcy​:icosahedron:

     in reply to

     • Woof?

     @Doggo (sorry that most of this is vague - while close, it's not 100% my field and all the big companies are very hush hush about how their systems _actually_ work)

   • halcy​:icosahedron: (halcy@icosahedron.website)'s status on Friday, 18-Jan-2019 04:03:26 JST halcy​:icosahedron:

     • Woof?

     @Doggo ah yes, unfortunately, https://www.youtube.com/watch?v=NMS2VnDveP8

     (this is slowly getting better, and so is speech recognition for Languages Other Than English, but US/GB English is still king wrt data amounts, far as I know)

   • halcy​:icosahedron: (halcy@icosahedron.website)'s status on Friday, 18-Jan-2019 04:21:46 JST halcy​:icosahedron:

     in reply to

     • Woof?

     @Doggo Hm, having tried it, maybe the pronunciation learning DOES work for recognition even in dictation mode (albeit it'll always assume that it's a name and it goes into places in sentences where names go).

     (how to: add name as a contact, get siri to speak the name [you can spell it letter by letter at her], tell her she's wrong and then teach her how to actually speak it in your voice. might take a few tries, but hey!)

   • halcy​:icosahedron: (halcy@icosahedron.website)'s status on Friday, 18-Jan-2019 04:30:52 JST halcy​:icosahedron:

     • Woof?

     @Doggo good luck!