are using it is a “very automated process”, Petro explained.
He said: “We’re always spidering out over the web to pick up new search terms that folks are using, which become part of the general models. There’s an ongoing natural adaptation that’s happening as the planet of human speakers using technology evolves on a day-to-day basis.
“We’ve also got commercial partnerships in terms of drug databases, a drug formulary of sorts, we keep track of the new drugs, we roll the names of those out into the models, and there’s an automatic updating process where the IT administrator on a specific site can roll those changes in.
them to get over the feeling that ‘wow, I’m not typing any more’. It’s a different workflow: they’re looking, they’re reading, they’re editing. But then once they get that, all of a sudden they’re off and running.”
Striving for 100%
Petro talked NHE through the recent development in speech recognition technology.
He said: “If you look historically at where we’ve come from and where we’re going, then 10 years ago speech recognition accuracy was probably bumping up against something like 75% accuracy, and from a medical perspective that’s obviously far too error-prone to be useful.
“Over the last 10 or 12 years, though, it’s gone from 75% into the high 90’s. If you look at radiologists for example, it’s not uncommon for them to have 98% or 99% accuracy all day long.
“In our research, we try to move that ‘edge’. For very clear speakers, they’re already topping out and we’re trying to get them from 98% to 98.5% accuracy, say, through algorithmic changes and noise filtering.”
Driving the technology
Improvements in technologies like noise filtering have also come to Nuance’s healthcare systems from the company’s other divisions,
such as those dealing with noise cancellation in cars.
He said: “We learn a lot from what happens in a car, because a car is very noisy, and you’re always speaking over tyre noise, wind noise, and that stuff finds its way into the speech recognition models.
“All the boats rise in the harbour, so to speak; as we add these things, everything gets better.”
He acknowledged that for heavily accented speakers and those who speak particularly unclearly, accuracy can be more like 91-92%.
He said: “For them, we put special programmes in place – gathering audio data, making adjustments to the models for low-volume speakers, for example. Then there’s research and investment going on in understanding it all – taking the speech, applying natural language processing to it, extracting clinically- orientated facts from it, and applying those facts downstream in analytics applications, clinical decision support applications, and that type of thing.
“We actually do very well with Indian-orientated accenting and Asian accenting. We develop specific acoustic models that are associated with a specific accent.
“Just small differences, such as between North American accents, can be accommodated for in those acoustic models. If you look at the
national health executive Jan/Feb 12 59
acoustic profile of each word – the WAV file, for example – there are material differences between one accent and another.”
Neologisms
Keeping the speech recognition software up-to-date with new medical terminology as people
“We try not to let a gap develop between the way people are speaking today, or new semantics getting introduced into the market, and the cutting edge of what our model supports.”
opinion@nationalhealthexecutive.com For more information
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56 |
Page 57 |
Page 58 |
Page 59 |
Page 60 |
Page 61 |
Page 62 |
Page 63 |
Page 64 |
Page 65 |
Page 66 |
Page 67 |
Page 68 |
Page 69 |
Page 70 |
Page 71 |
Page 72 |
Page 73 |
Page 74 |
Page 75 |
Page 76 |
Page 77 |
Page 78 |
Page 79 |
Page 80 |
Page 81 |
Page 82 |
Page 83 |
Page 84