Help train Mozilla's open source voice recognition

Help train Mozilla's open source voice recognition

Postby Julius » 18 Jul 2017, 15:23

See:
https://voice.mozilla.org/
very easy to use and potentially a very useful thing to have as an open-source model.
“You never change things by fighting the existing reality. To change something, build a new model that makes the existing model obsolete” - Buckminster Fuller
User avatar
Julius
Community Moderator
 
Posts: 1946
Joined: 06 Dec 2009, 14:02

Re: Help train Mozilla's open source voice recognition

Postby eugeneloza » 19 Jul 2017, 09:01

Fantastic!
However, I couldn't find neither feedback, nor my question answered in FAQ.
They do not list post-USSR countries in accent selection :) Neither preference in sound quality (I can record by both webcam audio and by semi-professional studio equipment). Also there are no guides for reading, e.g. should it be as neutral as possible, or with some artistic feel to it? Well, actually those are not too important, but might make the audition a bit better.

To everybody:
Recording your voice is not the only way to easily help the engine. You may just listen and mark other's auditions as valid, `cause it's damn hell of manual work.
User avatar
eugeneloza
 
Posts: 466
Joined: 22 Aug 2014, 12:15
Location: Ukraine

Re: Help train Mozilla's open source voice recognition

Postby Julius » 19 Jul 2017, 09:34

Yeah, the website is a bit focussed on English accents, while the reality is that the English voice recognition will also be used by a lot of non-native speakers. I wouldn’t worry about it too much, as in the end it will help the machine learning framework to better understand non-native voices. There are also already quite a few obviously non-native speakers if you listen to the sentences. Same goes for recording quality... it should in fact be non-perfect samples so that the system can better distinguish between background and recording noises and the actual voice. So by all means, use you web-cam and not your professional equipment.

However, what I am a bit concerned about is that they chose "book reading" examples. Somehow people have different and more intentionally pronounced voices when they read stories from a book, and this becomes very evident in the recorded samples. I am not an expert on voice recognition, but it would seem to me that this will somewhat distort the ML model towards this kind of task and that it will not work so well with the actual more common use-case of voice commands or transcription of natural conversations.
“You never change things by fighting the existing reality. To change something, build a new model that makes the existing model obsolete” - Buckminster Fuller
User avatar
Julius
Community Moderator
 
Posts: 1946
Joined: 06 Dec 2009, 14:02

Re: Help train Mozilla's open source voice recognition

Postby eugeneloza » 19 Jul 2017, 10:42

Yep. Noted that too.
User avatar
eugeneloza
 
Posts: 466
Joined: 22 Aug 2014, 12:15
Location: Ukraine

Re: Help train Mozilla's open source voice recognition

Postby Julius » 30 Nov 2017, 16:03

“You never change things by fighting the existing reality. To change something, build a new model that makes the existing model obsolete” - Buckminster Fuller
User avatar
Julius
Community Moderator
 
Posts: 1946
Joined: 06 Dec 2009, 14:02

Re: Help train Mozilla's open source voice recognition

Postby Lyberta » 01 Dec 2017, 10:17

Heh, I would like the ability to command bots in a video game using speech.
Some crazy person on the Internet.
User avatar
Lyberta
 
Posts: 312
Joined: 19 Jun 2013, 10:45

Who is online

Users browsing this forum: No registered users and 1 guest