Table 3.

Summary of articles including error analyses (n = 29)

Measure	Medical Domain	Articles	Summary of Study Designs and Findings
Percentage of documents with errors (n = 13)	Radiology	McGurk et al (2008)64	Study design Retrospective, cross-sectional by input method, with real reports24,41,64,66,85,87 Retrospective, cross-sectional by report type, with real reports24,85,86,90,132 Retrospective study with real reports84,88,89,91 Prospective, cross-sectional study with real reports24 Number of speakersa Median: 1986 Range: 241 to 14789 Number of documents evaluated Median: 30885 Range: 10024,41,91 to 584 87887 Percentage of finalized documents with errors Median: 26.9%86 Range: 4.8%64 to 71%91
		Pezzullo et al (2008)24
		Quint et al (2008)84
		Strahan and Schneider-Kolsky (2010)41
		Basma et al (2011)85
		Chang et al (2011)86
		Luetmer et al (2013)87
		Hawkins et al (2014)88
		du Toit et al (2015)66
		Ringler et al (2015)89
		Motyer et al (2016)90
	Emergency Department	Goss et al (2016)91
	Multiple	Zhou et al (2018)132
Mean errors per document (n = 7)	Radiology	Hawkins et al (2012)26	Study design Retrospective study of real reports88,90,91 Retrospective, cross-sectional by report type, with real reports132 Prospective, cross-sectional by report type, with real reports26 Controlled lab setting, cross-sectional by input method, with real reports44 Observational study, cross-sectional by input method, with real reports10 Number of speakersb Median: 1291 Range: 244 to 144132 Number of documents Median: 217132 Range: 2010 to 117326 Mean errors per document Median: 1.391 Range: 0.2490 to 2.544
		Hawkins et al (2014)88
		Motyer et al (2016)90
	Emergency Department	Zick and Olsen (2001)44
		Goss et al (2016)91
	Dentistry	Feldman and Stevens (1990)10
	Multiple	Zhou et al (2018)132
Accuracyc (n = 6)	Radiology	Herman (1995)92	Study design Controlled lab setting, cross-sectional by input method, with real reports20,44,94 Controlled lab setting with real reports92 Controlled lab setting with fictional patient scenarios93 Retrospective, cross-sectional study with real reports33 Number of speakersd Median: 1.5 Range: 120,93 to 533 Number of words evaluatede Median: 6019 Range: 727793 to 18 72192 Accuracy Median: 96.4% Range: 73%93 to 98.5%44, but often varied within studies based on the configuration of the SR system(s) evaluated
		Ramaswamy et al (2000)33
		Ichikawa et al (2007)20
	Emergency Department	Zick and Olsen (2001)44
	Nursing	Suominen and Ferraro (2013)93
	Unspecified	Zafar et al (1999)94
Word error ratef (n = 4)	Emergency Department	Zemmel et al (1996)59	Study design Controlled lab setting, cross-sectional by report type, with real reports19,95 Controlled lab setting, cross-sectional by SR system configuration59 Retrospective, cross-sectional by report type, with real reports132 Number of speakersg Median: 1295 Range: 759 to 144132 Number of documents Median: 46 Range: 759 to 217132 Number of wordsh Median: 60 874 Range: 11 56895 to 110 180132 Word error rate Median: 14.5% with general vocabularies, 11% with specialized vocabularies Range: 7.4%132 to 38.72%19 with general vocabularies; 5.21%19 to 9%59 with specialized vocabularies
	Internal Medicine	Devine et al (2000)95
	Otorhinolaryngology	Ilgner et al (2006)19
	Multiple	Zhou et al (2018)132
Other (n = 3)	Emergency Department	Hodgson et al (2017)57	Controlled lab setting, cross-sectional by input method; 35 participants were randomly allocated simple and complex clinical tasks 138 total errors with minor, moderate, or major potential for patient harm with SR across simple and complex tasks, vs 32 with keyboard and mouse
	Internal Medicine	Zafar et al (2004)96	Retrospective analysis of 148 real reports (104 created by 1 speaker with SR, 44 human transcribed with multiple speakers) 9 identified categories of SR errors, including enunciation, dictionary, suffix, added words, deleted words, homonym, spelling, nonsense, and critical errors
	Unspecified	McKoskey and Boley (2000)97	Unsupervised clustering of 1200 completed dictations from 6 speakers aligned with their original SR output Identified error clusters: short and function words; vowel destressing and cliticization; vowel syncope; words with sounds affected by telephony interference (eg, fricatives)

SR: speech recognition.

^a4 studies did not report the number of speakers.^66,87,88,90

^b4 studies did not report the number of speakers.^10,26,88,90

^cAccuracy = number of correctly recognized words/total number of words dictated.

^d2 studies did not report the number of speakers.^92,94

^e2 studies did not report the number of words evaluated.^44,94

^fWord error rate = (number of substitutions + number of insertions + number of deletions)/total number of words dictated.

^g1 study did not report the number of speakers.¹⁹

^h2 studies did not report the number of words evaluated.^19,59