Ask Alexa to do something for you when speaking in a strong Newcastle accent and she will often reply as if you were speaking Portuguese to her. Helpful intent but blank execution. It’s not a bug, exactly. It’s a much deeper structural issue and it’s been around for decades before any of these voice assistant products were invented.
Everyone Is Always Voting
Voice assistants are trained on large collections of recorded speech with transcripts (the actual words spoken in the recording). This data is then used to create a sound-to-word map, where the system becomes more confident in mapping a sound to a word the more it hears that sound mapped to that word. The system learns from clean interactions that add another data point to the map and treats failed interactions as noise that gets smoothed over.
So each and every utterance by a user in interaction with a voice assistant is effectively a vote on what to consider normal English. The way this works is described in Every Time You Do Something, You're Voting. And the people whose votes have counted most faithfully are the ones whose speech already matched the training set.
The training data was collected mostly in the United States and southern England, skewed toward younger RP-adjacent speakers. This wasn’t a deliberate decision to exclude other groups but rather because those were the early adopters, the paid volunteers, the people who happened to be in the room when the recording kits arrived. The machine learned to speak English as a new employee learns to operate in an office: from the first people to show up.
What a Confidence Threshold Actually Does
When you ask Alexa a question it calculates a probability of how likely that set of sounds will mean that particular word and if it’s above an internal confidence threshold it answers your question. If it’s below the confidence threshold it might ask you to repeat that question or get it wrong.
For those accents near to the ‘average’ of the data that was used to train Alexa, it works pretty well. But for those with strong Glaswegian, Geordie or Birmingham accents, sounds that are completely normal within one English dialect can fall below the confidence bar of the model. There are documented accuracy gaps of double figures for some of the strongest regional accents when compared to RP and estuary English speakers – not occasional misheard words but a consistent pattern on predictable sounds.
The machine isn't prejudiced in any human sense. It just treats unfamiliar input as probably wrong.
The Map That Fell Out
The failure rate correlates closely with the traditional North-South divide of accents and dialects in Britain, with strong Glasgow, Geordie, Scouse and Brummie accents failing to meet the confidence threshold more frequently, while rural accents fail more frequently than their urban counterparts. Older speakers fail more frequently than younger speakers.
Each time a speaker from one of those groups is misheard and gives up – such as putting down the device or ceasing to attempt to use it – this single failed interaction reinforces the system’s view that their speech is in some way an outlier. Those who are best served are those whose use most increases the quality of service. The rest are being quietly trained away from the norm.
—
Does this mean voice assistants are biased against regional accents?
The difference between the two communities occurred unintentionally, as no one engineer set up a system to be more effective in Surrey compared to Sunderland, but the result is still unequal. The simplest data to collect was from a set of communities and since then the feedback loop has widened the gap rather than closed it.
Can I improve how well Alexa understands my accent?
To a degree. Amazon and Google both offer personal voice profiles that improve recognition for individual users over time. The catch is that you have to persist through the early misreadings – which demands exactly the patience a frustrating first few days tends to use up.
Why haven't companies fixed this already?
Some have run targeted collection drives for underrepresented accents, and accuracy is slowly improving. But building a solid training corpus for strong regional speech is considerably harder than for RP, and because the devices already work well enough for most early adopters, the commercial urgency stays low.
Will regional accents eventually be treated the same?
Devices need to be used by enough speakers of different accents to generate sufficient training data. As it is, experiences are consistently patchy enough for the average user to cease use early on in the process, generating too thin a signal for any meaningful reduction in the ‘gap’ to occur. Someone needs to go out of their way and deliberately break this cycle of underuse.
FAQ
Why does Alexa struggle with strong regional accents?
Alexa’s sound-to-word matching is done by a model trained on the bulk of its speech data from the United States and southern England, skewed toward younger RP-adjacent speakers. Sounds from a strong accent that fall below the internal confidence threshold of the system are assumed to probably be wrong, and therefore the user is asked to repeat themselves or it is otherwise misunderstood by the system.
Which accents do voice assistants have the most trouble with?
Strong Glasgow, Geordie, Scouse and Brummie accents fail to meet the system’s confidence threshold most often. Typically the pattern follows the traditional North-South divide, with rural as opposed to urban accents, and older as opposed to younger users, failing more often. Some regional accents show a double-figure difference in accuracy compared to RP and estuary English.
Can I improve how well Alexa understands my accent?
Yes, to a degree. Amazon and Google allow users to set up a voice profile, which improves recognition for the individual user over time. Persistence through the early misreads is required, and these can be frustrating in the first few days of set-up, using up the patience the process needs.
Why haven’t companies fixed accent bias in voice assistants?
Yes, some targeted collection efforts for underrepresented accents have been undertaken and some improvements in accuracy have been made. However, building a robust training corpus for strong regional speech is much harder than for RP, and as devices already work well enough for most early adopters of speech recognition, there is little commercial urgency to try to improve matters further.
