[Resubmitting again to Show HN ... previously I linked to a blog post about the game, but this time I linked to the game directly, according the Show HN rules]
We're planning to do a follow up blog post with some analysis of the data -- comparing the interesting results.
We'd be happy to share the data with anybody that is interested. All in all, there's the sounds clips + the game data + the results data -- I think it all needs to be considered together in order to draw any good conclusions.
If you are a researcher or are just interested in what comes out of the game, we'd love to compare notes with you! (And I'm sure we can stick some sort of license onto the data if that's useful.)
We're using commercially available speed-to-text platforms (Watson, Google, Microsoft, etc.), so we're just evaluating how well they are doing with this test. I think all of those are using NN's internally -- but since we're doing a cross-platform test, it's probably not useful for training the individual platforms.
With much larger traffic and fancier gameplay, I think this technique could be used to generate a higher volume of training data on a custom built speech-to-text system (eg. Kaldi)
10 comments
[ 2.2 ms ] story [ 23.2 ms ] threadHere's the companion blog post:
https://www.linkedin.com/pulse/we-made-game-robots-round-rob...
The Signl.fm crew is here to answer any questions!
We'd be happy to share the data with anybody that is interested. All in all, there's the sounds clips + the game data + the results data -- I think it all needs to be considered together in order to draw any good conclusions.
If you are a researcher or are just interested in what comes out of the game, we'd love to compare notes with you! (And I'm sure we can stick some sort of license onto the data if that's useful.)
We're using commercially available speed-to-text platforms (Watson, Google, Microsoft, etc.), so we're just evaluating how well they are doing with this test. I think all of those are using NN's internally -- but since we're doing a cross-platform test, it's probably not useful for training the individual platforms.
With much larger traffic and fancier gameplay, I think this technique could be used to generate a higher volume of training data on a custom built speech-to-text system (eg. Kaldi)
https://medium.com/@TheMindWright/helping-robots-learn-to-sp...
jk totally played the game lol