Every tool you point at callers is also pointed at your agents
The pitch is caller verification. The exposure nobody mentions is the other direction: a genuine agent, on a narrowband line, from a home connection, flagged as synthetic — and a workflow that turns that flag into a conversation about their job.
What you are actually dealing with
Contact centres see synthetic speech from three directions at once, and they need separating because they have different answers.
Impersonated customers. A cloned voice with harvested personal data, calling to change an address, reset a credential or authorise something. If you run voice biometrics on the inbound leg, this is an attack against a control you already depend on.
Automated callers at volume. Synthetic voices used to probe your IVR, harvest confirmations or run enumeration attacks. The signal here is pattern and volume more than any single verdict.
Your own audio, misread. The one your operations director will hear about first. Agents work on compressed lines with aggressive noise suppression, often from home, sometimes with a headset that colours the signal. Every one of those conditions moves a genuine recording towards the reading that also indicates synthesis.
We work on recordings, not on live legs. If the requirement is a real-time decision while the caller is still on the line, that is an inline product and a different vendor. What we do is take the call from your recording platform afterwards and give you a documented answer on it.
The reverse exposure, in detail
Suppose you wire an analysis into QA sampling. It runs across a slice of calls each week. One morning it returns a synthetic verdict at high confidence on a call handled by an agent with four years’ service.
What happens next is entirely a function of a process you wrote before this page loaded. In an organisation with no policy, the flag reaches a team leader, who reads it as evidence that something dishonest happened, because that is what “synthetic detected” sounds like. The agent is asked to explain. They cannot. There is nothing an innocent person can produce that proves a recording of their own voice is real — the burden is unmeetable by design. Their explanation is “that was me”, which is exactly what a guilty person would say.
The likely truth is duller: their broadband degraded, the codec renegotiated mid-call, noise suppression stripped the room. But by the time anyone establishes that, the agent has sat through an investigatory meeting over an allegation nobody can name, and the trust cost is already paid. That is the real risk of deploying this badly, and it is borne by someone earning a fraction of what the software costs.
So decide these before you switch anything on. Who sees a flag first, and is it someone technical rather than someone managerial. Whether audio conditions are shown alongside the verdict every time, because they are usually the explanation. Whether a flag can ever route automatically into a performance or disciplinary queue — it should not. And whether the agent is told, which they should be, along with the report and the error rate.
What this cannot do for you
It cannot screen a live call. Recordings only. No inline verdict, no agent prompt, no real-time hold.
It cannot tell you who is speaking. No speaker matching, no identity confirmation, no view on whether the caller is your customer. That is a biometric question.
It cannot separate two speakers on one mixed recording. If agent and caller are on a single mixed channel, vocal-tract consistency cannot be attributed to either and you will get unclear. Send stereo or per-leg recordings if your platform can produce them.
It cannot compensate for narrowband audio. Your recordings have passed through codecs designed to discard the detail detection depends on. Confidence is reduced accordingly, and unclear will be a common and correct answer on your estate.
It cannot be the reason anyone is disciplined. A named human weighs the analysis with everything else and owns the decision. A verdict that automatically becomes an outcome is a system that will eventually end someone’s employment over a codec renegotiation.
A deployment that will not hurt anyone
Cases, not everyone
Start with disputed calls and confirmed fraud cases. Bulk screening of every agent and every customer is a much larger processing activity and needs its own justification.
Flags go to fraud, not to line management
Fraud analysts read audio conditions as context. Team leaders read “synthetic” as an accusation. Route accordingly and log every flag with its confidence and quality ceiling.
Write it into policy first
Agents and customers should know the analysis exists, what a flag does and does not mean, and how to challenge one. Introducing it silently is what makes a later case unwinnable.
Recording, monitoring and analysing calls involving both customers and employees carries obligations that differ by jurisdiction, including consent to recording, worker consultation and transparency duties. [VERIFY: verify the rules in each country you operate sites in] This is not legal advice.
Operational questions
What audio should we send?
The highest-quality copy your recording platform holds, per-leg if possible, and ten to thirty seconds of one person speaking cleanly rather than the whole call. A long mixed recording with hold music and crosstalk will usually return unclear.
Can this run against our QA sample automatically?
The API can. Whether it should is the question this page is about. If you do it, treat the output as a research signal held by a small team, not as an alert visible to operations, until you have measured your own false positive rate on known-genuine agent calls.
How do we baseline it on our own audio?
Run a few hundred calls you already know are genuine, across your worst lines and your remote agents, before you trust a single flag. If the rate on that set is uncomfortable, that is the finding — and it is better to have it in a spreadsheet than in an employment tribunal.
Does an agent have a right to see the result?
In most places, where the result is used in a decision about them, yes in substance. Plan to disclose the report, the engine version and the error rate as a matter of course rather than on request.
Baseline before you deploy
Measure it against calls you know are real, on your own lines, before it sees a single case.
Reviewed