Slashdot Mirror
Is Speech Recognition Finally 'Good Enough'?
jcatcw writes "Speech recognition software is fast, but it still may not be accurate enough. Clerical jobs usually ask for 40 wpm, but speech recognition software can keep up with someone speaking at 160 wpm. In Lamont Wood's demo it did very well at too/two/to and which/witch, but will it still render 'I really admire your analysis' as "I really admire urinalysis'? At 95% accuracy, people aren't jumping on the bandwagon. Wood's typing speed is about 60 wpm with 93% accuracy, so he found that using speech recognition was about twice as fast as typing. Those who type at hunt-and-peck speeds will experience results that are even more dramatic. There's really only one product on the US market: Dragon NaturallySpeaking from Nuance Communications. The free versions from Microsoft aren't up to the task and IBM sold ViaVoice to Nuance, where it's treated as an entry-level product."
Dear aunt, let's set so double the killer delete select all.
Well, there's spam egg sausage and spam, that's not got much spam in it.
I'm using Dragon NaturallySpeaking. Right now, as I write this calm it, comet, post, and it sure as hacking beats typing.
Actually, I am using Dragon NaturallySpeaking right now, and it works very well. It actually works better if you speak quickly (as you normally would) and it's pretty good at inserting grammar along the way. I have bilateral tendinitis, and the software has been a godsend for me. I was even able to finish writing my book, a task that was becoming just too painful typing manually.
Oh, and you are probably wondering how long it takes to train the software? About a half an hour, and I find the accuracy at around 95%.
SEO Copywriter. Just Say ON
It all depends what sort of corpus the SR system is trained on. So yeah, foreigners will have problems because a system trained for, say, British English will not perform well with American English. For this same reason an SR system trained for "normal" speech will do very poorly with lyrics in music.
As for stuff like "i really admire your analysis" being interpreted as "i really admire urinalysis," that stuff can easily be ironed out by an n-gram based system that "ranks" English sentences based on probability. What is the chance that "urinalysis" will follow "your" which follows "admire"? Such things can be estimated well enough if you use a large corpus to train your n-gram system (as long as the corpus you're using for this is the same "kind" as whatever speech the SR system is interpreting -- that is, newswire, business meeting, etc.)
I like basketball!!1!
I am presently a financial customer of an enterprise speech recognition product that Nuance offers. For several years now, the speech recognition software industry has been under consolidation, with Nuance buying a few different competitors and technologies. Most recently, this dance has continued with Nuance being acquired by ScanSoft, a company known for specializing in type recognition.
Nuance support is marginal at best, and through all the consolidations, understanding even within their own company of how the product works is quite lacking. We have found our own developers often times educating the Nuance support folks in various aspects of how the product is working, and then inquiring as to whether this is intended behavior or not. Crickets can often be heard finishing these types of conversations. We normally would have moved to another product under these conditions, but simply put - Nuance acquired what little was left, and now has no competition in the market. Competition is what spurs innovation, and so with the continued consolidation, it is hard to see significant advances in the technology without free help from academia.
If you think the Microsoft monopoly is bad, imagine if they absorbed Apple and somehow took over Linux leaving you with a few "choices", but all under the Microsoft moniker. The technology is very neat and the enterprise level products do some basic things quite well, but there is still some glaring room for innovation that I don't expect anytime soon under present industry conditions.
n-gram based language models are nothing new. Statistics is all fun and dandy, but it's no panacea. It might just be enough to throw in an even larger corpus (something like the complete Google index), but it's still hard. (BTW, n-gram Markov chains more or less originated in speech recognition, to get the individual phonemes right, and I'm quite sure they're doing at least something like it at the word level these days. It still sucks, as the quality users demand for proper dictation is extremely high.)