Peer-reviewed research
Racial disparities in automated speech recognition
Allison Koenecke et al., Proceedings of the National Academy of Sciences, 2020 · DOI 10.1073/pnas.1915768117
Read the originalRead on October 16, 2026
- What it is
- A study of five commercial ASR systems from Amazon, Apple, Google, IBM and Microsoft, transcribing structured interviews with 42 white speakers and 73 Black speakers from five US cities.
- How much weight it carries
- Peer-reviewed and widely cited. The corpus is 19.8 hours of interview audio, not dictation. The results show engine differences by speaker group in 2020 systems, not current dictation apps.
Facts taken from this source
- In the study, Apple's system had a word error rate of 0.45 for Black speakers and 0.23 for white speakers.
“the WERs for black and white speakers are 0.45 (SE: 0.005) and 0.23 (SE: 0.003)” (Results)
- The authors trace the speaker-group gap to the acoustic models used by the ASR systems.
“We trace these disparities to the underlying acoustic models used by the ASR systems” (Abstract)
- In the study, Microsoft's system had a word error rate of 0.27 for Black speakers and 0.15 for white speakers.
“the WER for black speakers is 0.27 (SE: 0.004) compared with 0.15 (SE: 0.003) for white speakers” (Results)
- The 2020 study found that error rates for Black speakers were nearly twice as large as for white speakers in every case.
“the error rates for black speakers are nearly twice as large in every case” (Results)
- Across five commercial ASR systems, the average word error rate was 0.35 for Black speakers and 0.19 for white speakers.
“with an average word error rate (WER) of 0.35 for black speakers compared with 0.19 for white speakers” (Abstract)
- Researchers describe word error rate as a standard measure of the discrepancy between machine and human transcriptions.
“a standard measure of discrepancy between machine and human transcriptions” (Methods, Word error rate)
- The study defines word error rate with the total number of words in the correct transcript as its denominator.
“is the total number of words in the ground truth” (Methods, Word error rate (text after Equation 1))
Pages that rest on this source
- Voice Assistants vs Dictation: Siri, Alexa, GeminiGuide
- Why is Apple Dictation so bad?Question
- Speech Recognition: What It Is and How Well It WorksTopic
- Dictation Languages, Accents and Voice TranslationTopic
- Google Docs Voice Typing: How to Use It and Its LimitsGuide
- Speech Recognition and Accents: Accuracy and FixesGuide
- Dictation Mistakes and How to Make Voice Typing AccurateGuide
- Is speech recognition less accurate for some people?Question
- Why doesn't voice recognition understand my accent?Question
- Acoustic modelGlossary
- Dictation and Voice Typing: How to Talk Instead of TypeTopic
- Dictation Software: How to Choose the Right AppTopic
- Speech to Text for Accessibility, RSI and DisabilitiesTopic
- Transcription: How to Turn Audio and Video Into TextTopic
- What Is Dictation Software? Types and How It WorksGuide
- What Is Dictation? Voice Typing and Talk to Text ExplainedGuide
- Best Dictation Software and Apps, Compared by NeedGuide
- How Accurate Is Speech Recognition? The Real NumbersGuide
- Transcription Software: How to Choose OneGuide
- How to Transcribe an Interview: Audio to TextGuide
- Dictation for Academics, Researchers and ThesesGuide
- How accurate is voice recognition software?Question
- How do researchers dictate field notes?Question
- How do you improve dictation accuracy?Question