Home » Services » Examples of adjustments and discrimination: service providers » Fraud detection through voice analysis

Fraud detection through voice analysis

Disclaimer – please read
This page does not apply outside Great Britain.
Last updated 26th May 2020 (part update 26th October, 2023).

Voice risk analysis (or “voice stress analysis” or “lie detector”) technology, perhaps using AI, is sometimes used in telephone calls on insurance claims, and by some local authorities. It aims to spot people who are more likely to be giving false information. This may unfairly disadvantage people who stammer.


  • These voice analysis systems may potentially flag up things resulting stammering, whether the stammer is overt or hidden.
  • It is suggested that organisations should not use output of a voice analysis system if they know or suspect the caller has a stammer, or indeed another disability which is liable to affect the output.
  • Workers using such systems should be trained in recognising possible stammering and other relevant disabilities, and how they may be mistaken as dishonesty.
  • With the customer’s consent, an organisation could flag a customer’s communication disability on their system. In any event, where voice analysis is being used, organisations could consider asking callers if they have a stammer or other relevant disability.
  • In deciding under the Equality Act whether use of a voice analysis system is justified, one issue is likely to be whether the organisation can show there is sufficient evidence that their system is effective in reducing fraud generally.

STAMMA (the British Stammering Association) suggests:

“Fraud detection
AI or human fraud detection systems may misidentify stammering as hesitancy, being unsure or lying. Find out whether your fraud detection systems use speech dysfluency or pauses as a marker for fraud. Use customer flagging procedures to override this particular marker for customers who stammer.”
STAMMA Stammering & customer contact guide, 2023. However note that service providers’ EqA obligations are not likely to depend simply on whether the customer has previously asked or agreed that the provider flag the stammer on its system.

About voice analysis

Different systems may vary, but broadly as I understand it: At the start of an interview the software is “calibrated” by asking basic questions such as name, address and date of birth. This is taken as “normal”, and the software looks at changes in voice patterns when further questions are asked. It shows a Low Risk message if the voice pattern matches the baseline. However if certain parameters are breached it shows a High Risk message, indicating there may be a lie. The results can be taken into account in deciding whether a claim should be fast-tracked or subjected to more rigorous scrutiny.

A call centre worker may be told to use the voice analysis result only as one indicator among others (though this may be of limited use if the worker mistakes effects of stammering as dishonesty), and/or not to use it where he or she knows the person has a relevant disability such as a stammer. However presumably different organisations will have different policies, and different implementation in practice by call centre workers.

There is more in an article on the STAMMA (British Stammering Association) website: Worries over voice risk analysis to combat crime (stamma.org), from summer 2006.

Equality Act implications and stammering

How far people who stammer are affected by voice analysis will depend on the system, but they may potentially be flagged as High Risk because of things resulting from their stammer. Depending on the software, it may for example pick up hesitations produced by the stammer. Or even if the person sounds fluent, it may pick up effects of stress from trying not to stammer, effectively from trying to “deceive” people that they are fluent. See below Problems using voice analysis with stammering, and article on the British Stammering Association website: Worries over voice risk analysis to combat crime (stamma.org).

Detriments to the person who stammers from being flagged as high risk will depend on the circumstances, but may include:

  • being subjected to extra questioning (which a person who stammers may find particularly difficult on the phone) or perhaps to other checks,
  • perhaps delay in receiving a payment to which the person is entitled (or not receiving it?),
  • depending on the system, perhaps it being on the person’s record that they were flagged as high risk, even though they were later found to be honest; and
  • potentially the offensive stigmatising nature (see CHEZ) of being put though extra checks because of the stammer.

S.15 EqA

If the service provider or public body using voice analysis is aware of the stammer (most obviously if the customer has told them of the stammer, or is having some struggles in speaking), they may need to disregard any negative results from voice analysis. Failure to do so may be contrary to s.15 EqA as discrimination arising from disability, on the basis that the person who stammers would otherwise be treated unfavourably without sufficient justification. The software’s results will (presumably) not be useful in such cases, or at least its effectiveness with stammering will probably not have been tested, so discrimination is likely to be unjustified. There should not be a problem if (a) interviewers are told to disregard High Risk results from people who stammer (or other relevant disabilities) and (b) the interviewer realises the person has a stammer (or other disability) and does indeed follow the instructions.

However there is also potential Equality Act liability under s.15 even if the service provider does not know of the stammer, if it fails to show that it could not reasonably have been expected to know of the disability. Partly for this reason (and also to justify use of voice analysis and to make any reasonable adjustments) it may be an obligation under the Equality Act to adequately train interviewers to recognise – so far as possible – a person who has a stammer (or other relevant disability) even where it is not obvious. A person who stammers will often try and hide it as far as possible. The stammer may be having effects on speech which are not clearly stammering: see below Problems using voice risk analysis with stammering, and Disability: Hiding the stammer>How people hide.

The service provider might consider whether to ask a caller if they have a speech impairment or other disability which might affect the voice analysis. This raises various issues however. For example a significant number of people who have a stammer may be reluctant to say so, GDPR issues (including on health data) need to be considered, and a fraudulent caller might potentially answering “yes” when they didn’t have an impairment.

Quite apart from voice analysis issues, the service provider etc may be open to challenge under the Equality Act, eg s.15, if it does not train staff in how effects of stammering (and potentially other disabilities) may be mistaken as dishonesty. Otherwise it may eg treat customers unfavourably through staff disbelieving them as a result of things which are actually due to their disability.

Reasonable adjustments and indirect discrimination

Even where the service provider or public body does not know of the stammer (or other disability) and cannot reasonably be expected to know of it, there is potential Equality Act liability for indirect discrimination and failure to make reasonable adjustments. In the case of service providers and bodies exercising public functions, there is no explicit defence here as regards not knowing about the disability (see Reasonable adjustments by service providers>Knowledge of disability and Knowledge of disability>Indirect discrimination – is knowledge required?).

In looking at whether indirect discrimination against people with speech impairments or other relevant disabilities is justified, one issue is likely to be whether the service provider or public body can show there is sufficient evidence that the voice analysis system they use is effective in reducing fraud generally.

Also, to show justification or to show that further reasonable adjustments were not required, the service provider may need to have taken steps such as staff training in recognising relevant disabilities (see above) so that adverse impact on disabled people is minimised.

Insurance claims

Some (perhaps many) insurance companies use voice analysis software on people phoning up to make claims, to help flag up any that are likely to be dishonest. Claims flagged up by the software are more likely to be subject to particularly rigorous investigation. See Worries over voice risk analysis to combat crime (stamma.org), summer 2006. As outlined above, these arrangements may involve a breach of the Equality Act.

Benefit claims

Voice analysis was piloted for welfare benefit claims from 2007. Representations were made to the Department for Work Pensions (DWP) about the problems of using the software with disabled claimants. As a result of the trials, in autumn 2010 the DWP abandoned plans to introduce voice analysis for welfare benefits, saying it was not good value for money. However it was reported in 2014 (link below) that some local authorities were still using voice analysis, or considering doing so. Links:

As well as possible breach of Equality Act rules on public functions, there may be a breach of the public sector equality duty (PSED) unless stammering and other relevant disabilities are properly considered.

Problems using voice analysis with stammering

  • “Basic” questions such as name and address are often a particular struggle for someone who stammers. Yet these basic questions are apparently used to “calibrate” the software to the interviewee’s “normal” voice pattern, enabling it to detect any change on the later more probing questions. Where a person struggles – or is more stressed – on those basic questions, one would not expect the software to give useful results.
  • Different voice analysis systems may measure different things, but may well view as suspicious things which are actually results of stammering, whether the stammer is open or hidden. These might be for example:
    • Hesitations, repetitions, filler phrases such as “well”, “you see”, if it measures these.
    • Changes in voice from stress or strain related to stammering, or stress related to trying not to stammer. Voice analysis software aims to spot deceit, and many people who stammer effectively try to “deceive” people that they are fluent, to try to appear normal. This is what might be called a ‘symptom’ of their stammer.
  • If the person is notified when their speech is going to be analysed by a lie detector, this may serve to increase stress levels and dysfluency, which may make them more likely to be suspected of fraud.
  • The interviewer may have been told to disregard software results for eg someone who stammers, but a key feature of stammering is that the person often tries to hide the stammer. The interviewer may not realise they have a stammer.
  • Presumably the software has not been tested – and therefore not validated – with stammering. Also different people stammer differently.
  • See also Mistaking stammering for dishonesty (in context of court appearances).

For issues of using voice analysis where a person has a disability more generally, see archived Disability Alliance Factsheet: Voice risk analysis and benefit claims (archive of disabilityalliance.org), 2011.

20th anniversary of stammeringlaw, 1999-2019