Speech recognition: what the research says
Quinlan, T. (2004). Speech Recognition Technology and Students With Writing Difficulties: Improving Fluency. Journal of Educational Psychology, 96(2), 337.
This quantitative research study, published in a peer-reviewed academic journal, investigated how composing with Speech Recognition (SR) and advance planning affected the writing performance of less fluent writers. Using a between-subjects, repeated measures design, the study addressed the question of whether low-load transcription and planning support resulted in improved narratives for children with writing difficulties. A total of 41 children, ages 11 to 14 years, participated in the study. Two groups of children, consisting of fluent and less fluent writers, composed a series of four narratives under four different writing conditions which included two levels of transcription (i.e., SR and handwriting) and two levels of planning (i.e., with and without advance planning). All students received training in advance planning and the use of SR software. Dragon Naturally Speaking Professional 5.0 (2002) was the SR program used. Statistical analysis, a repeated measures analysis of covariance (ANCOVA), was used to determine whether writing mode and planning might influence the written products of fluent and less fluent young writers.
RESULTS: Narratives handwritten by less fluent writers were rated significantly inferior to the handwritten narratives of fluent writers in terms of length, quality, and surface errors. The narratives composed by less fluent writers using SR had significantly more words and fewer errors than their handwritten narratives. In contrast, SR did not improve fluency or accuracy of fluent writers. After advance planning, both groups of children spent more time composing. Advance planning may facilitate an attentional shift to text generation, whereas SR may increase the efficiency of text generation by reducing transcription-related interference. It is important to note the less fluent writers in this study represented a select subgroup of children whose writing difficulties were primarily transcription related. Children whose writing difficulties derive from expressive language problems may not experience the same gains.
MacArthur, C. A., & Cavalier, A. R. (2004). Dictation and Speech Recognition Technology as Test Accommodations. Exceptional Children, 71(1), 43-58.
This research study, published in a peer–reviewed academic journal, investigated:
- Can high school students with and without disabilities learn to use speech recognition to produce text with acceptable accuracy?
- What are the effects of dictation to a scribe and dictation via speech recognition technology, compared to handwriting, on the writing test performance of students with and without learning disabilities?
The study used a repeated measures group design. The participants included 31 high school students: 21 students with LD that affected their writing and ten students without LD. All students received training in the use of speech recognition for dictation then wrote essays under each of the following three conditions:
- handwriting (HW),
- dictation to a human scribe (DS), and
- dictation to a computer using the speech recognition software Dragon Naturally Speaking, Version 4 (SR).
Both groups of students were able to quickly learn to use speech recognition to produce text with acceptable accuracy. Two-thirds of the students achieved 85% accuracy and more than one-third achieved 90% accuracy. Only three were below 80% accuracy. No statistically significant differences in accuracy on sentence or word probes were found between LD and NLD students. Repeated measures analysis of variance with two between-group measures (disability status and gender) and one within-group measure (writing condition: HW vs. DS vs. SR) was conducted on the overall quality scores.
RESULTS: The results demonstrate that both dictation conditions did help students with LD produce better essays. However, composing condition did not affect vocabulary. The best essays were produced when dictating to a scribe. In this condition, students were free to concentrate on the content, organization, and wording of their essays without concern for mechanics. Essays that were produced by students with LD when dictating to speech recognition software were not as high in quality as when using a scribe, but were better in quality than their handwritten essays. In this condition, students did not need to think about spelling and handwriting, but they had new cognitive burdens, such as speaking clearly and monitoring their writing for errors. In comparison to handwriting, the effect size for dictation to a scribe was large and that for speech recognition was moderate. Both dictation supports were educationally as well as statistically significant. No statistically significant differences among conditions in quality of writing were found for students without LD.
Higgins, E. L. & Raskind, M. H. (1995). Compensatory Effectiveness of Speech Recognition on the Written Composition Performance of Postsecondary Students with Learning Disabilities. Learning Disability Quarterly, 8(2), 159-174.
This pioneering research study on speech recognition software, published in a peer-reviewed academic journal, investigated the effectiveness of speech recognition supported writing compared to writing without assistance and the more traditional strategy of using human transcribers. The participants included 29 university students previously identified as having a learning disability with a specific disability in written language. All students received training in the use of speech recognition for dictation then wrote essays under each of the following three conditions: (a) using the speech recognition system Dragon Dictate, Version 1.01 (SR); (b) dictating the essay to a human transcriber (TR); and (c) without assistance (NA). Students were allowed to handwrite or use a word processor without the spell-checking function to generate their ‘no assistance’ essay. Handwritten and transcribed essays were typed verbatim. All 81 essays were each scored by two readers and given a single holistic score using criteria from the Upper Division Written Proficiency Exams. Three essays received different scores between the readers. A tie-breaker reader scored these three essays and agreed with one of the initial readers to determine the score. All readers were blind to the condition under which the essays had been written. Inter-rater reliability between readers was .93.
RESULTS: The results demonstrate a statistically significant difference when comparing the holistic scores of the SR essays with the NA essays. Data provided by the exploratory analysis of the compositions suggest a possible reason for this difference. Statistical analysis through a stepwise multiple regression found the most sensitive predictor of the holistic score was ‘big words’; words of seven or more letters. Interestingly, throughout the study many students with spelling difficulties mentioned a typical writing strategy they used was to substitute a ‘baby word’ they knew they could spell for the word they really wanted to use to avoid the embarrassment of spelling it incorrectly.