Why voice assistants mishear accents, and what helps.

How recognition goes wrong

A speech recogniser turns sound into words using models trained on many hours of recorded speech (how speech recognition works). Researchers score it with the word error rate: the share of words it gets wrong, leaves out or adds. A rate of 0.19 means about one word in five is wrong; 0.35, about one in three. Voices that sound different from the training recordings, in accent, dialect, age or the way sounds are formed, tend to get more errors.

What studies found

• Five commercial systems, United States. Allison Koenecke and colleagues ran structured interviews with 42 white and 73 Black speakers in five US cities, about 20 hours of audio matched on age and gender, through five state-of-the-art systems. All five showed substantial gaps: an average word error rate of 0.35 for Black speakers against 0.19 for white speakers. The gap was just as large on identical phrases, which points to the acoustic models rather than to word choice.
• A Dutch system. Siyuan Feng and colleagues measured the bias of a state-of-the-art Dutch recogniser by gender, age, regional accent and non-native accent, and traced the errors to differences in how sounds are articulated.

Both studies propose the same remedy: train on more varied voices, and measure performance for each group, not just on average.

What you can do

1. Get close and get quiet. A microphone near your mouth, away from noise, helps every voice.
2. Short sentences: one request at a time.
3. Check before it acts. Read what it heard before a message goes or a meeting moves.
4. Prefer assistants that confirm. A misheard word in a question costs a second; in a sent message, more: an AI that asks first.

What makers should do

• Test with many accents, ages and dialects, and publish error rates per group.
• Train on varied voices, including non-native speakers.
• Design so that a mishearing cannot turn into an action without the user seeing it.

How we approach it

For Lychnos Ring, understanding speech comes first: if the ring cannot understand you, nothing else matters. We test speech recognition through a simulated ring microphone in six languages, with quiet rooms, noise and mumbled speech: can a ring understand a mumble? The ring is designed to show the first line of what it understood and to wait for one press before it sends or changes anything. It is in development.

Sources, read 4 November 2026: Koenecke and colleagues, racial disparities in automated speech recognition, PNAS, 2020; Feng, Kudina, Halpern and Scharenborg, quantifying bias in automatic speech recognition, Interspeech, 2021. Abstracts read at Europe PMC and arXiv.