4 ms·
Speech recognition has been ok for a long time if you happened to have an accent that worked fine with it. I don't. I speak English with a relatively broad Sca
by vidarh 10y ago
Speech recognition has been ok for a long time if you happened to have an accent that worked fine with it.
I don't. I speak English with a relatively broad Scandinavian accent. Up until a few years ago, no voice recognition system I tested managed to recognise what I said with a high enough precision to be usable.
This "suddenly" changed over the last few years. And it may very well be that the approaches changed (e.g. more heavy emphasis on statistics and large scale feeding of voice data for bad recognition to people to help improve the models; offloading to data centres), but it did change for me at least. E.g. I can now not just have it recognise valid English words, but Youtube will recognise channel names that are mixes of words and letters/numbers, and Amazon Fire will recognise every actor name or movie title I trow at it with a precision rate I would not have expected just a 3-4 years ago.
It's to the point where the Google Docs speech recognition is almost usable for me too.
- YeGoblynQueenne 10y agoThe fact that voice recognition now understands your accented English is not a sign of its increased ability to understand different accents: only of its increased coverage of specific accents which now include yours. Not sure if I'm explaining that well enough. What I mean is, say Siri had a 90% accuracy in English before; that would be because it was trained with a lot of samples of English of many different accents. It wasn't trained with enough Scandinavian accents so it didn't understand them very well. Eventually more Scandinavian accented English examples became available and Siri's accuracy with Scandinavian accented English improved as a consequence. But that doesn't mean that Siri is good at understanding English with any novel accent. So its improvement is not an improvement in terms of quality, only in terms of quantity: it can do more of what it did before, but nothing new. Which is not to say that speech recognition doesn't work or that small improvements like that are useless, or anything like that, it's just to say there are still limitations and unless you're very careful it's easy to misunderstand what those are.