The Accent and Dialect Accuracy Gap

Automatic speech recognition does not fail randomly. It fails more for speakers whose speech is less represented in training data, and that is a documented, measured and persistent finding.

What the research shows

Published studies comparing commercial speech recognition across speaker groups have repeatedly found substantially higher error rates for some groups than others, with the largest documented gaps falling along racial and regional lines in US English.

The gap has narrowed with newer models and it has not closed.

It is worse for: speakers of regional and ethnic dialects, non-native speakers, older speakers, speakers with atypical speech, and speakers in noisy environments — which correlates with income.

It is a data problem more than an algorithm problem. Models learn from what they were trained on, and training corpora historically over-represented some speakers.

Why it matters in a contact centre

Two distinct harms, and they land on different people.

On agents. If agent speech is transcribed less accurately for some agents, then any automated scoring, compliance checking or coaching signal derived from transcripts is systematically less accurate for them. An agent whose required disclosure is transcribed wrongly fails a compliance check they passed.

This is an employment matter, not a technical curiosity. Automated evaluation feeding performance management, applied unequally by accent, is a discrimination exposure.

On customers. If customer speech is transcribed less accurately, then topic detection, sentiment measurement and escalation flagging work less well for those customers. Their complaints are less likely to be surfaced, their distress less likely to be detected.

The aggregate effect is a system that sees some customers more clearly than others and reports the result as an objective picture.

Measuring it in your deployment

This is testable and almost nobody tests it.

Stratify your accuracy sample by speaker group, using whatever grouping is meaningful and lawful for you — region, first language, queue serving a particular population.

Compute word error rate per group.

Compare. A gap of several percentage points is common. A large gap means any downstream automated judgement is unequal.

Do the same for your critical terms, which is where compliance failures would land.

Repeat after any model change, because vendors update models and the profile changes.

What to do about a gap you find

Do not use automated scoring for compliance or performance where the accuracy gap is material, until it is addressed. This is the single most important consequence.

Keep humans in the loop for anything with an employment or a customer outcome.

Improve the inputs: better audio, custom vocabulary, model adaptation on your own speakers.

Raise it with the vendor, with your measurements. Ask what their per-group accuracy is and whether they test for it. Many do now; the answer tells you how seriously they take it.

Document the finding and the mitigation. If an automated process is later challenged, evidence that you measured the disparity and responded is a materially different position from having never looked.

The transparency question

Agents should be told how automated evaluation works and what its limits are, which is covered separately.

An agent who suspects the system mishears them is frequently right, and a programme with no route to raise that produces justified resentment and no correction.

A dispute mechanism that can pull the audio and compare it to the transcript resolves this, and it also generates the data that shows whether the problem is systematic.

The general principle

Any automated measurement applied to people should be checked for whether it works equally well across the people it is applied to.

For speech recognition this is not hypothetical or precautionary. The disparity has been measured repeatedly, it is known to the vendors, and a deployment that does not test for it in its own environment is choosing not to know.

Presenting the finding internally

A measured accuracy disparity is an uncomfortable finding and how it is framed determines whether it is acted on.

Lead with the measurement, not the implication. "Word error rate is X for group A and Y for group B on our own sample of thirty calls."

State what it affects. Which automated checks depend on transcription, and therefore which determinations are less reliable for which agents.

State the immediate mitigation. Suspending automated scoring for the affected items, or requiring human confirmation.

State the durable fix. Audio improvement, vocabulary, model adaptation, vendor engagement.

Frame it as a control failure rather than a scandal, because that is what it is: a measurement instrument that works unevenly, discovered by the programme's own quality checks.

A programme that finds and reports this is functioning. One that has never looked has the same problem and no evidence of it.