[Edeka does all--machine speech recognition in social medicine expert testimony].

UNLABELLED

Automatic speech recognition systems are already being used in spheres employing a restricted vocabulary.

OBJECTIVE

Our aim was to investigate whether low-cost speech recognition software for PC is capable of being usefully employed in the sphere of sociomedicine.

MATERIALS AND METHODS

To this end 34 representative pages of text (a total of 11,000 words) taken from expertises on cases of suspected medical malpractice (many different subspecialties) were dictated using IBM's "Voice Type Simply Speaking" software. Having completed a page, the resulting error rate was recorded, and the text was corrected before we proceeded with the dictation. Finally, 3 pages of text were re-dictated and the resulting error rate determined.

RESULTS

The error rate in the previously unknown text ranged between 10 and 23 per cent (mean 15.9%) without any significant reduction during the training phase, while that in the re-dictated text was drastically reduced to less than 3 per cent. It became evident that once a word was corrected the system hardly ever repeated that particular mistake.

CONCLUSION

The system's poor performance on unknown text and the missing reduction in the error rate during the training phase are obviously not due to any incompetence of the system but to the huge amount of technical jargon in the scope of medical writing. To attain an acceptable performance we suggest to either extend the training phase, or, preferably, to confine the application to a single medical subspecialty. Its overwhelming learning ability makes the system a serious candidate typist in the sphere of sociomedicine.