Search PubMed⌕ Search
Home/Archive/PubMed

PubMed · 8862666

Text processing by digital voice recognition.

G A Meijer·J P Baak·P J van Diest·A H van Hattum·H C van der Linden·J J Koevoets
View original source ↗Export metadata ↓

Abstract

OBJECTIVE: To evaluate digital voice recognition in free text mode. STUDY DESIGN: A standard medical text (182 words and punctuation marks) was dictated four times each by six people using a commercially available voice recognition system. RESULTS: The time used, including the time needed for making corrections, decreased from a median of 14.5 (6.3-18.8) minutes in the first session to a median of 6.7 (5.1-15.7) in the fourth session. the number of corrections necessary in the last session was a median of 20.5 (15-97)-i.e., 11% of the number of words in the text. All users described working with the system as highly strenuous. Dictating the same text with a dictaphone took median 1.4 (0.9-1.6) minutes and about five minutes of typing and correction time. CONCLUSION: Although the total time was about the same with the two methods, the time for the pathologist increased fourfold when using the voice recognition system. To obtain acceptable input speed, an appropriate template module is essential.

Explore related subjects

Signal Processing, Computer-AssistedTime FactorsVoiceJournal Article

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

G A Meijer, J P Baak, P J van Diest, A H van Hattum, H C van der Linden, J J Koevoets. 1996. Text processing by digital voice recognition.. https://pubmed.ncbi.nlm.nih.gov/8862666/

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related citations

PubMed1999

[Modular PC-based data acquisition and processing system for biological signals].

The data acquisition system described here is designed for biomedical research and permits the recording of up to eight biological signals simultaneously. A personal computer using the Windows 95 operating system is employed for data monitoring, data processing and analysis during experiments. The system has been designed for reliability, economy, flexibility and ease of handling, with the aim of achieving universal application. To avoid interface incompatibility, problems with transfer protocols and the data formats of commercially available products, analog signals are used for further processing. The individual input channels are electrically isolated from one another and the PC to avoid ground loops, and for reasons of safety. An isolated voltage supply is available for pre-amplifiers and bridges. A bandwidth of 0-5 kHz and the maximum sampling rate of 12.5 kHz suffice to pick up higher frequency signals such as EMG and ENG. The modular software and hardware concepts permit the use of almost any desktop or laptop PC as a central processing unit. The PC handless documentation, data acquisition, data analysis and the preparation of publications. If needed, further analytical functions can be added in modular form. Finally, the option of saving data in the ASCII format permits processing of results with such standard software packages as Excel, Access, Matlab and Origin.

Signal Processing, Computer-Assisted↗
PubMed1998

A quantitative method for the detection of edges in noisy time-series.

A modification of the edge detector of Chung & Kennedy is proposed in which the output provides confidence limits for the presence or absence of sharp edges (steps) in the input waveform. Their switching method with forward and backward averaging windows is retained, but the output approximates an ideal output function equal to the difference in these averages divided by the standard deviation of the noise. Steps are associated with peak output above a pre-set threshold. Formulae for the efficiency and reliability of this ideal detector are derived for input waveforms with Gaussian white noise and sharp edges, and serve as benchmarks for the switching edge detector. Efficiency is kept high if the threshold is a fixed fraction of the step size of interest relative to noise, and reliability is improved by increasing the window width W to reduce false output. For different steps sizes D, the window width for fixed efficiency and reliability scales as 1/D2. Versions with weighted averaging (flat, ramp, triangular) or median averaging but the same window width perform similarly. Binned above-threshold output is used to predict the locations and signs of detected steps, and simulations show that efficiency and reliability are close to ideal. Location times are accurate to order square root of W. Short pulses generate reduced output if the number of data points in the pulse is less than W. They are optimally detected by choosing W as above and collecting data at a rate such that the pulse contains approximately W data points. A Fortran program is supplied.

Signal Processing, Computer-Assisted↗