Speech Recognition Technology - Finally Ready for Prime Time?

With the release of the iPhone 4S and Siri, Apple has introduced speech recognition (SR) technology to the masses. Apple bills and markets Siri as a “humble personal assistant.” However, I doubt many radiologists, who have been working with SR technology for multiple years, would describe their SR software systems this way.

Radiologists who use SR software to dictate their reports often describe the experience as frustrating. Much of this frustration, I believe, stems from the introduction of SR before its maturity.

Apple is renowned for their elegant consumer gadgets. They delay release of their products until their engineers believe that a technology has improved to a point where its implementation will not cause undue frustration on the part of their users. Case in point, while the tablet computer existed long before the iPad, Jobs and Co. at Apple realized that it would not be readily embraced by the average user until the product could be very thin, very responsive, and light. Apparently the “magic” (to borrow a term Jobs liked to throw around) year when technology was robust enough to accomplish this was 2010.

Now that Apple has determined SR technology to be mature enough to incorporate it into its iconic iPhone, should radiologists stop their grumbling?

Speech recognition software, which facilitates radiology dictations, has many proven benefits. Last year in AJR I wrote about our experience at UNC Hospitals incorporating SR software into the radiologist’s workflow. We noticed substantial improvements in report turn around time similar to what other observers had demonstrated, but also found that improvements in turn around time varied significantly between users due predominately to different work habits. Users who spent greater time training the system to recognize their voice and adding and revising terms in the vocabulary library, for example, benefited the most.

If you read between the lines of Apple’s marketing onslaught you will realize that they advise the same type of behavior to improve Siri (the more you use it, the better it will get). Additional benefits of SR driven dictation systems include the ability to leverage templates, mini macros, and structured reporting to improve accuracy, consistency, and billing. Moreover, advances in natural language processing coupled to SR software may allow for the ability to integrate targeted radiology decision support at the time of dictation, the ability to structure unstructured reports, and identify key content and highlight critical findings.

However, despite these benefits SR software has potential drawbacks. Most notably errors can result from incorrect capture of a radiologist’s voice and more important, the intent of their language. Most recently Sarah Basma and her group at Women’s College Hospital in Toronto, found an alarmingly high rate of errors using SR software vis a vis traditional transcriptionists for breast imaging reports. After correcting for multiple variables Basma’s group found that reports generated with SR were eight times more likely to contain major errors than those generated by traditional transcription.

The conflicting reports on SR software should give us pause as we move forward. The companies producing SR software have improved their products significantly over the last few years. Despite these improvements, a recent Diagnostic Imaging poll shows that while 80 percent of users use SR software, 30 percent still find it hard to use and not as accurate as traditional transcription services.

Apple believes the technology is ready for prime time. However as in most things tech, what is good enough for the average consumer product doesn’t necessarily stand up to the rigorous standards of healthcare. If Siri misinterprets your request for the weather forecast, you may get rained on. However, if a radiologist’s breast imaging report is recorded incorrectly the potential consequences are much more substantial.