4 ms·
> Google Now has a pretty wide range of behaviors, but it's the spooky-accurate voice recognition that sets it clearly ahead. In the last year, Google Now has o
by kuschku 11y ago
> Google Now has a pretty wide range of behaviors, but it's the spooky-accurate voice recognition that sets it clearly ahead. In the last year, Google Now has only misunderstood me ONCE.
> She also mishears me 25-45% of the time, AND she seems to be getting worse over time. I assume this is because they are trying to widen her audience beyond a standard California accent, but it's incredibly frustrating for a product to get worse, not better, over time. I now use her only for voice dictation of texts on my iPhone, and use Google Now in app form exclusively for more advanced queries.
And this is pretty much a huge issue for this market. Entry requires having already spent several billions on voice tech.
You’d have to practically use all of YouTube as training set. Which Google did.
Maybe someday we can have laws that require all neural networks trained with data from others to be in the public domain, even if the people providing training data agreed to the data being used for training.
Then we could compete on other things than "understands voice".
- lgas 11y agoWell, until then, Google's speech recognition API is available (https://cloud.google.com/speech/ https://cloud.google.com/speech/) so it seems like it should be possible to at least compete at Google levels of SR without even having to understand what training data is.
- kuschku 11y agoAnd how exactly am I supposed to do that without transferring userdata to a third party (like Google)? Privacy laws – which I personally like – prevent using that.
- shostack 11y ago"You’d have to practically use all of YouTube as training set. Which Google did." Didn't they use Google Voice as well?
- Touche 11y agoWait. So Google recognizes a long time ago that machine learning would be important to their business and so invests in a variety of projects to aid with that; Google 411, Google Books, etc. But because Apple either didn't have the foresight or chose not to invest in these technologies, Google should be forced to donate their data sets? To one of the richest companies in the world no less? throws up hands
- kuschku 11y ago> But because Apple either didn't have the foresight or chose not to invest in these technologies, Google should be forced to donate their data sets? To one of the richest companies in the world no less? No, I’m not arguing for Apple. I don’t care – they can afford it. I, as a student, can’t afford it. You, if you wish to start such a startup, can’t afford it. And you know what? Google couldn’t do any of this on its own. Instead, they tricked people into working for Google for free. And now they try to make others pay for the work of volunteers that transcribed YouTube videos, that solved Captchas, etc. headdesk
- ttam 11y agohow exactly is it "working for google for free" when people got to host their videos for free on google's platform... there's a reason why you don't directly pay for these services, you know? > I, as a student, can’t afford it. You, if you wish to start such a startup, can’t afford it. well, if I wish to start such a startup, I would probably work on an alternative e.g. technology that could overcome the need for all that data..
- dharma1 11y agoI don't think it will take billions. LSTM+CTC speech recognition implementations are coming to OSS and I think we have large enough available data sets for training without needing Google size resources. Audio books are a good resource for training and it's not hard to amass hundreds or thousands of hours of training material that way. I think we will have 95%+ accuracy ASR as open source trained networks very soon. In the case of Amazon Echo/Dot they benefit greatly from the far field array mic for isolating a clean speech signal in noisy/far field environments, would be nice to see generic USB array mics with Linux drivers
- kuschku 11y agoThat’s indeed a good thing. With audiobooks – maybe rented from the local library? – and public TV we could have quality voice recognition tools. Hopefully also ones in German, French, etc. Currently only English is well supported by the non-cloud ones, which is quite annoying.
- beagle3 11y agoWhat's the OSS LSTM+CTC package that's coming along?
- dharma1 11y agohttps://github.com/srvk/eesen https://github.com/srvk/eesen I expect there will be something in Kaldi soon, too