Detect cloned voices. Analyse phone deception.
VoiceGuard is BlueWave AI’s protection against deception on phone calls. It gives separate signals on three questions: could the voice be synthetic or cloned, does the conversation show a deceptive pattern, and can the call’s origin be verified. The three capabilities are at different stages and complement each other.
The outlined button opens an email.
Conversation analysis: performance validation underway. Runs today in our own development and demo environment; no customer production deployment yet. Designed for customer-controlled processing, with deployment options agreed for each implementation.
Three questions,
three capabilities
Voices are cheap to clone. Phone fraud also works with a real voice.
A cloned voice can ask for money, data or an approval in the voice of a relative, a manager or a member of staff. But the attack is a chain of deception that also works with a real human voice: urgency, posing as someone else, demands for secrecy, pressure to skip the usual security step. That is why VoiceGuard does not stop at asking whether the voice is real.
Voice: could it be synthetic or cloned?
VoiceGuard Engine. Works on live audio streams and recorded files, including narrowband telephone audio. Several consecutive segments are assessed together to reduce false alerts from brief noise. Each alert can be traced to the model version and the segments behind it.
It does not identify the speaker and is not legal evidence.
Conversation: does it show a deceptive pattern?
Vishing and social engineering pattern analysis. Flags deception even when the voice is real; in our demo it also flags calls made with a real human voice.
It does not read minds and is not a lie detector.
Origin: can the caller be verified?
BlueWave Witness. The calling organisation signs its own calls, and the person receiving the call sees one of three states (below). A proof of concept runs on a simulated telephone channel.
It is not a finished mobile app.
The signals are not merged into a single “fraud / not fraud” verdict. The customer’s or partner’s own rules decide the next step: proceed, verify further, suspicious, block or inconclusive.
The call really comes from the organisation. This verifies its origin, not that everything said is true or that every request is safe.
No signature. On its own, a missing verification is not a sign of fraud; confirm any money movement on another channel.
No signature, and detection also fires. Hang up and call back on a known number. This is not a legal finding of fraud.
Witness complements detection: detection asks what may be fake, Witness asks what can be verified. Status: planned partner capability, with a proof of concept on a simulated telephone channel.
CallCenter Guard
CallCenter Guard: VoiceGuard in the contact centre.
For banks, insurers, outsourced contact centres, helpdesks and any team that handles requests by phone. It checks calls where a fake voice could get through, starting with the recordings you already have.
Access and password resets
A caller asks to recover an account or reset a password. Is the voice synthetic, and is the conversation following a pressure script?
Customer data changes
New device, new phone number, changed contact details: requests where one convincing call can redirect everything that follows.
Sensitive phone approvals
Phone approvals and callbacks for actions that matter, where urgency and secrecy are the usual levers.
Internal helpdesk
A caller claiming to be a colleague asks IT for access or a reset.
Executive or supplier impersonation
A call in the voice of a manager or a supplier asks for a payment or an approval now.
Review after the fact
Reviewing suspected fraud calls afterwards, on recordings.
These situations are illustrative, not customer cases. We do not claim a separate, validated solution for each of them.
- Whether the signal is useful in one selected process.
- At what false-alert rate, and with what processing time.
- What share of the audio is unusable, and what review workload results.
These are measures, not promised results.
- The live call connection (SBC/SIPREC) is in development.
- Bringing alerts into the agent or fraud-management workflow is planned, and built together with the customer.
- CallCenter Guard does not identify the speaker, prove fraud or end a call. Your agents stay in charge.
How it connects
Built as an observation branch: it analyses a copy and returns signals.
VoiceGuard does not block, hold or route the call. Today the audio it works on is your recordings; live inbound calls follow once the connection is built.
Partner integrations
Built to be integrated into other people’s services.
Partner integrations
We work with partners to integrate VoiceGuard capabilities into their existing platforms and services. Technically this uses a REST/WebSocket API and events; the production security layer is part of the rollout.
Through a partner to end users (B2B2C)
Banks, operators or app providers can offer protection to their own customers under their own brand. A planned mobile service, CallGuard Mobile, is part of this route.
A mobile app does not automatically get access to the audio of a phone call; the route to the audio is designed with each partner, platform by platform. This is a possible cooperation model, not an existing agreement.
The technology
What sets the technology apart.
Control over data
Processing can run in your own environment; an on-premise installation package is in development. Customer data is not used for model training by default.
Adaptable to your data
The trainable layer is small, so it can be fine-tuned on approved customer data. The engine is built on Kolmogorov–Arnold Network (KAN) layers.
Built for telephone audio
It works on narrowband audio. Capacity and accuracy are measured on your hardware and data rather than quoted in general.
Traceable results
Each alert can be traced to the model version and the segments behind it, so a reviewer can see what the signal rests on.
The call audio is cut into short segments.
The VoiceGuard Engine assesses each segment.
One result for every segment.
Several consecutive segments are assessed together.
Reduces false alerts from brief noise or distortion.
The exact number of segments is deliberately not published: it would tell an attacker what to stay under.
Context
External sources
Context: what scams already cost.
Regulator and survey figures, for background. They describe scams in general, not VoiceGuard results, and they are not specific to phone fraud.
First assessment
The first joint assessment.
- A short call about one process or user scenario, and the audio that is lawfully available for it.
- An assessment on existing recordings or approved samples, with a written report and a joint decision on whether and how to continue.
- We agree the scope, terms and any fees of an assessment before it starts.
We have presented our voice detection to teams at three Hungarian banks.
How licensing is structured
For contact-centre use we propose an annual licence on the concurrent channel capacity you commit to, rather than per-minute pricing. The cost then follows the capacity you plan for, not every call you decide to cover.
Capacity and accuracy are measured on your hardware and data first.
The outlined button opens an email to office@bluewaveai.co. The booking button opens a Google Calendar appointment page.