Heidi Horstmann Koester

Model simulations of user performance with word prediction

Previous work has demonstrated that, under well-defined conditions, a quantitative model can accurately represent user performance with word prediction systems (Koester, 1994; Koester & Levine, 1994, 1995, 1997). This paper illustrates the use of this model to simulate user performance across a broad range of conditions. Examples of simulation results are presented to show how user characteristics, strategy of use, and system characteristics combine to determine overall performance. The use of model simulations to inform clinical decision-making and research questions is discussed.

User Performance With Speech Recognition: A Literature Review

The application of speech recognition to the computer access needs of people with disabilities continues to grow, and a greater understanding of user performance with such systems is needed. This article reviews what is known about user performance with speech recognition systems, with a focus on its application to accommodation of physical disability. Although current systems offer the potential of text entry at 150 words per minute, the literature suggests that users actually achieve somewhere between 8 and 30 words per minute. Barriers that may contribute to this gap, such as the costs associated with correcting recognition errors, are reviewed, and directions for future research are proposed. A major need is for additional research involving users who have physical disabilities.

Research in computer access assessment and intervention.

Computer access technology (CAT) allows people who have trouble using a standard computer keyboard, mouse, or monitor to access a computer. CAT is critical for enhancing the educational and vocational opportunities of people with disabilities. Choosing the most appropriate CAT is a collaborative decision-making process involving the consumer, clinician(s), and third party payers. The challenges involved and potential technological solutions are discussed.

Selecting an Appropriate Scan Rate: The “.65 Rule”

Investigators have discovered that the ratio between a users reaction time and an appropriate scan rate for that user is approximately .65, which we refer to as “the .65 rule.” As part of a larger effort to develop software that automatically adapts the configuration of switch access software, data were collected comparing subject performance with a scan rate chosen using the .65 rule and a scan rate chosen by the user. Analysis of the data indicates that for many people, the .65 rule produces a scan rate that is approximately the same as the average switch press time plus 2 standard deviations. Further analysis demonstrates a relationship between the coefficient of variation (the standard deviation divided by the mean) and error rate. If accurate information is available about the mean, standard deviation, and distribution of a clients switch press time, a scan rate can be chosen that will yield a specific error level. If a rigorous statistical approach is impractical, the .65 rule will generally yield a usable scan rate based on mean press time alone.

Toward automatic adjustment of keyboard settings for people with physical impairments

Purpose. We are developing a software system called IDA (Input Device Agent), whose goal is to optimally configure input devices for people with physical impairments. This study assessed IDAs ability to recommend three keyboard parameters in response to measurements of typing performance: repeat rate, repeat delay, and use of StickyKeys. Method. Twelve typists with physical impairments participated. The study employed a repeated measures design. Each participant typed six sentences in each of four keyboard conditions: default settings, IDA-recommended repeat settings, StickyKeys On, and a repeat of default settings. Results. Two participants had significant problems with inadvertent key repeats, when using the default repeat settings. For those two participants, use of the IDA-recommended repeat settings reduced the number of repeated characters by 96% and significantly improved text entry rate and typing accuracy. IDA recommended StickyKeys for six participants, each of whom had at least one problem related to modifying keys without StickyKeys. Use of StickyKeys for these individuals eliminated their modifier-related errors and significantly improved typing speed. IDA did not recommend StickyKeys for the six participants who demonstrated no need for it. Conclusions. The results indicate that IDA can provide useful assistance with repeat settings and StickyKeys.

Modeling One-Switch Row-Column Scanning with Errors and Error Correction Methods

Single switch scanning has lots of different configuration options. One way to choose the most appropriate configuration for a client is to use a model to predict performance under different configurations. Most existing models expect error-free performance, however, and none integrates all the types of errors that can occur and the variety of error- correction methods that are available. A model is presented which predicts user performance for single-switch row- column scanning with errors and error-correction methods. The model is used to draw conclusions about the utility of different error-correction methods.

Validating a model of row-column scanning.

Purpose: For individuals with severe motor and communicative disabilities, single switch scanning provides a way to access a computer and communicate. A model was developed that utilizes scanning interface settings, error tendencies, error correction strategies, and the matrix configuration to predict a user’s communication rate. Method: Five individuals who use single switch scanning transcribed sentences using an on-screen keyboard configured with the settings from their communication devices. Data from these trials were used as input to a model that predicted TER for the baseline configuration and at least three other system configurations. Participants transcribed text with each of these new configurations and the predicted TER was compared to the actual TER. Results: Results showed that predicted TER was accurate to within 90% on average. The scan rate was also entered into a previously published model which assumes error-free performance. For our model, the average error for each participant was 10.49%, compared to 79.7% for the model assuming error-free performance. Conclusions: Our model of row–column scanning was much more accurate than a model that did not consider the likelihood of an error occurring. There is still room for improvement, however, and the results of the study will lead to additional modifications of the model. Implications for Rehabilitation Although row-column scanning is a very slow method of selection, changes in the configuration of the interface can produce noticeable changes in performance. When configuring a row-column scanning interface, clinicians should consider the type of errors their client is likely to commit to target interface features to their clients specific needs. Some clients who use row-column scanning may not benefit from advanced interface features, even if they are available.

Method for enhancing text entry rate with single-switch scanning.

Single-switch scanning is a technique used by some individuals for spoken and/or written communication. We developed a method for adjusting the settings in a single-switch scanning interface to increase a users text entry rate (TER). We evaluated that method with nine individuals who use single-switch scanning to communicate. Text entry rates improved by an average of 120% (p = 0.003). All nine subjects increased their TER by at least 40%, and five of the nine increased their TER by over 100%. At baseline, TER averaged 1.42 words per minute (wpm), ranging from 0.28 to 2.92 wpm. With the revised settings, TER averaged 2.72 wpm and ranged from 1.12 to 6.51 wpm.

A model of performance cost versus benefit for augmentative communication systems

Many augmentative and altemative communication (AAC) systems provide a benefit of keysavings at the c a t of increased cognitive-perceptual requirements. A user performance model is derived that shows how cost and benefit determine text generation rate in these systems. The accuracy of the model is evaluated, and examples of useful applications are presented. Introduction Augmentative and altemative communication (AAC) systems can provide people who have disabilities with a means of independent communication in the area^ of speech, writing, and computer applications. but a major problem continues to be the slow speed of text generation. Attempts to address this through new interface designs have focused on utilizing the users motor abilities as efficiently as possible. While these designs can improve motor efficiency, they may also place increased cognitive and perceptual requirements on the user, lading to unknown effects on overall performance. Our goal is to improve the undersmding of how this tradeoff affects text generation rate in AAC systems, with a current focus on word prediction systems. These systems attempt to predict an intended word based on previously chosen letters. Word choices are displayed in a short list and refined as the user selects letters. Since many words can be completed by choosing from the List rather than single letter spelling, the number of required selections can be substantially reduced. However, overall text generation rate may not improve [1,2,3.4], because the time required to make each selection may increase. due U) additional cognitiveperceptual processes such as searching the word list [2.5,6]. Our primary approach is to develop and validate quantitative models that can simulate and predict user performance under different system and user characteristics. While our previous efforts in this area have focused on specific system implementations [5,6], this paper presents a more general model that has application to a wide range of AAC interface designs and input methods. The derivation of the model is presented using word prediction as the example, but the principles and the equations are identical for other techniques such as abbreviation expansion or symbolic encoding. Methods A useful way to consider the effect of word prediction on text generation rate is as a balance between cost and benefit. Cost results from the extra selection time involved in using word prediction, and is defined as the decrease in the rate at which items (i.e.. letters and words) can be selected. Its mathematical exDression is:

Toward automatic adjustment of pointing device configuration to accommodate physical impairment

Purpose. Software was developed which makes recommendations regarding configuration of a computer pointing device, such as a mouse, to accommodate a persons physical impairment. Specifically, a software agent automatically recommends a setting for the computers control-display gain based on observations of a users performance in a target selection task. Method. The software agent makes its recommendations based on available adjustment settings in the existing operating system. The agent was evaluated in studies with 12 participants who have motor impairments. Results. The agent-selected gain was not associated with significant improvements in selection time or error-free performance compared with the operating systems default gain. Across participants and trials, gain did not have a significant effect on selection time except at the lowest gain settings tested. However, two participants did have notable and consistent improvement in selection time and error-free performance using the agent-selected gain; gain across trials had a significant effect on number of target entries and number of submovements; and a post-hoc analysis indicated improved target selection time when varying both target size and control-display gain. Conclusion. These observations provide possible avenues for future work, although the current study indicates that changes to control-display gain, alone, are unlikely to offer improvements in speed or accuracy for the general population of people with motor impairments.