Electronic Letters to:

The Practice of Informatics: Comparative Evaluation of Three Continuous Speech Recognition Software Packages in the Generation of Medical Reports Eric G. Devine, Stephan A. Gaehde, and Arthur C. Curtis J Am Med Inform Assoc 2000; 7: 462-468 [Abstract] [Full text] [PDF]

eLetters: Submit a response to this article

Electronic letters published:

Is Minimal Training sufficiently predictive of the accuracy levels achievable in the Real World?

Gowthaman G   (18 March 2002)

Higher Speech Recognition Accuracy with Language Models

James M Maisel   (30 October 2001)

Were environmental factors studied?

Joseph Foss   (7 February 2001)

Is Minimal Training sufficiently predictive of the accuracy levels achievable in the Real World? 18 March 2002
Gowthaman G, Medical Officer BBR Hospitals, Hyderabad, India Send letter to journal: Re: Is Minimal Training sufficiently predictive of the accuracy levels achievable in the Real World? gg{at}jipmer.net Gowthaman G	Sir, With respect to the article in JAMIA {7:462-468 (2000)}: Comparative Evaluation of Three Continuous Speech Recognition Software Packages in the Generation of Medical Reports, I have the following observations to make- 1. The physicians in the study had only completed the minimal training necessary for the Speech Recognition Software. It must be noted that different software packages have different requirements for what they consider the minimal necessary user inputs before the software begins to work on the voice inputs to decipher speech and convert it into textual material. The greater the amount of training the program gets, the better the accuracy. Assuming safely that a physician in real practice who will use the software to enter Patient’s Records eventually will complete the full training as needed by the software in question, the comparative accuracy of the softwares may differ substantially from that gotten in the study. Thus the study, though indicative, may not truly reflect on the accuracy levels that may be gotten in practice. Conclusion: In future Studies on the accuracy of the Speech Recognition Software, full training of the software would be a better option, and this is what we should aim for. 2. This one is for the Software Manufacturers: The Speech Recognition System analyzes the speech in real time and finds out the most probable word from its inbuilt Thesaurus. The user has one final look at the text and make changes in it before he/she saves the final draft of, say the Case Record. If the software were to track the changes that the user were to make before saving the document, then the software would be able to ‘know’ the nuances and eccentricities in the user’s speech patterns and be able to respond better in the future to the same or similar sounding words. This would increase the accuracy of the software over time. While this will also increase the computational power needed for the functioning of the software or conversely decrease the speed of recognition of language, and might not be a viable option at present CPU speed, it certainly is something to work on given that CPU speeds are constantly getting better. For now, it may be incorporated in a Speech Recognition System thus: whenever the system has a doubt (say <90% accuracy) it may silently ‘mark’ the word after having made its ‘guesses’ and once the final draft is complete and checked by the user, the system will ‘check’ its own accuracy and incorporate this data in its database, so that in the future the program would be able to deal with the voice more accurately. Dr. Gowthaman G. Medical Officer BBR Hospitals Hyderabad, India
Higher Speech Recognition Accuracy with Language Models 30 October 2001
James M Maisel, Ophthalmology NYU School of Medicine Send letter to journal: Re: Higher Speech Recognition Accuracy with Language Models jmaisel{at}zydoc.com James M Maisel	All of the tested speech recognition engines work on 3 levels: phonetic, vocabulary and language models. Recognition is greatly enhanced by including a statistical language model designed for word combinations by specialty. End users can only add words easily and lack the computation linguistics knowledge needed to enhance these products. The accuracy of our products are over 97% out of the box with minimal training. The products work well even with foreign accents but good dictation habits are the most important variable. Noise cancelling microphones make the products work in any environment and can also be used with Sony or Olympus handheld recorders with excellent results. ZyDoc is currently looking for doctors for a University Sponsored recognition accuracy project in all specialties. Jim Maisel, MD President, ZyDoc.com 1455 Veterans Memorial Hwy Hauppauge, NY 11788 631-273-1963 jmaisel@zydoc.com
Were environmental factors studied? 7 February 2001
Joseph Foss, md University of Chicago Send letter to journal: Re: Were environmental factors studied? j-foss{at}uchicago.edu Joseph Foss	I started with the IBM systems, but found the acoustics of the room (as documented with the software) may require retraining for different environments. It would be helpful to know the type of room that these trials were done in and whether there was any thought given to comparing different areas. It would be less practical to have to have different profiles for each area I might dictate in. thanks