Generator traces across languages
A language-disjoint evaluation of forensic representations. Train on 12 languages, then test the same generators on six unseen languages.
91.0%Top-1 generator retrieval on languages excluded from training
Featured in TechCrunch · Read the storyAudio authentication models.
Built for real conversations.
We build models that detect synthetic speech and verify speaker identity. Use them across calls, recordings and voice agents, in the infrastructure you control.
- Latency per clip
- 24msPer clip · vendor-reported
- Detection accuracy
- 94.47%Scored by Podonos
- Telephone channel
- 8kHzModel design target
As featured in
Press and recognition
Ecosystem support
Programmes and ecosystem partnersA few seconds of audio can clone a customer’s voice.
A familiar voice can now be generated from a short recording. Banks and contact centers need to verify the speech behind an account change, payment or identity check.
Account takeover through the contact center
- Attack
- A caller uses a cloned voice to pass voice authentication or convince an agent, then resets credentials or changes the registered number.
- Control
- Aletheia scores the customer leg before any account change.
Payment authorisation and impersonation
- Attack
- A cloned executive, relative or customer asks for an urgent transfer, or confirms a high-value payment by phone.
- Control
- High-risk instructions from a flagged voice are stepped up or blocked.
Remote onboarding and identity verification
- Attack
- Synthetic or replayed speech is used to open accounts or pass liveness prompts during remote identity verification.
- Control
- Deepfake detection runs alongside Eidos and liveness checks.
#1 in commercial latency.
24 ms per clip, with 97.3% of deepfakes detected. Results from the Podonos benchmark across 23 systems.
- Latency per clip
- 24ms#1 among commercial systems
- Processing speed
- 2.3×the next-fastest commercial system
- Deepfakes detected
- 97.3%of synthetic audio in the test
- Files scored
- 4,524Every file. None declined.
- Detection accuracy
- 94.47%#9 of 23 systems
Speed and detection accuracy
Higher is more accurate. Further right is faster.
Detectif.aiOther commercial systemsAccurate and fast: 90%+ accuracy, 10×+ real time
Processing speed · 1× = real time
94.47% accuracyat 119× real-time speed
Commercial latency ranking
Average time per audio clip. Lower is faster.
- 1Detectif.ai24 ms
- 2Pella Research57 ms
- 3NII Synthetiq Audio91 ms
- 4Corsound AI180 ms
- 5Pindrop282 ms
- 6Resemble DETECT-World399 ms
- 7Fennura553 ms
- 8Hive881 ms
- 9Whispeak1.1 s
- 10Resemble AI1.2 s
- 11Reality Defender5.7 s
Reported timings · bars use a log scale
Accuracy scored by Podonos. Detectif.ai timings are vendor-reported; hardware and service conditions vary. The latency ranking includes the 11 commercial systems with published timing.
Podonos source · September 2026How the models read a phone call.
Follow one cloned-voice call from the raw line signal to a decision. Every stage below is drawn from the same few seconds of audio.
Capture the call audio
The customer leg arrives as 8,000 samples a second, compressed with µ-law. Detection starts from this signal, exactly as the network delivers it.
- Input
- 8 kHz mono · µ-law
Our models.
Four specialized models for speech authenticity, speaker identity, call classification and conversation.
Aletheia · Truth
Deepfake detection
Analyze live calls and recordings for synthetic, cloned or replayed speech. Return a synthetic score and identify the segments that require review.
- Synthetic speech
- Voice cloning
- Replay detection
- Input
- Live or recorded audio
- Output
- Synthetic score + flagged segments
AuthenticityExample
Flagged speech
Synthetic score0.93
Synthetic speech detected
Applications
- Live call screening
- Recording analysis
- Remote onboarding
Integrations.
Stream live calls from your telephony platform, add signals to a voice agent, score recordings, or run the detector on the handset itself.
Fork the customer leg from your telephony stack.
Send the caller’s audio as it arrives and act on the signal in your own routing, without changing what the agent hears.
- SIP and RTP media forks
- Contact-center audio streams
- Agent desktop alerts
stream-call.js
// Stream the customer leg of a live call
const session = await detectif.calls.stream({
callId: "7F3A-21C9",
audio: { encoding: "mulaw", sampleRate: 8000 },
checks: ["authenticity", "speaker"],
speakerReference: customer.voiceprintId,
});
customerLeg.on("audio", (frame) => session.send(frame));
session.on("signal", (signal) => {
if (signal.authenticity.synthetic >= 0.5) {
fraudDesk.route(call, signal); // your policy
}
});
Signal returnedapplication/json
{
"callId": "7F3A-21C9",
"at": 6.85,
"authenticity": { "synthetic": 0.92, "segments": [[5.50, 10.08]] },
"speaker": { "match": 0.86 }
}Example syntax. Production interfaces are provided during evaluation.
Voice agents.
Classify the recipient, verify the caller and run the approved workflow. Transfer the call and its context when a specialist is needed.
- Noesis
Classify the recipient
Identify a person, IVR, voicemail or another agent and select the appropriate route.
- Aletheia + Eidos
Check the caller
Inspect authenticity and compare the enrolled voice before a sensitive request.
- Workflow rules
Apply access controls
Check permitted actions, tool access and additional verification requirements.
- Lexis
Act or hand off
Continue in scope, or pass the conversation and the evidence to your team.
ConversationExample
- NoesisHuman recipient · conversation route selected
- Lexis
- Caller
- Access rulesNew payee · additional verification required
- AletheiaSynthetic score 0.93 · voice flagged
- EidosSpeaker match 0.86 · authenticity check failed
- Lexis
- OutcomePayment held · call and evidence sent to a specialist
Deployment options.
Use the hosted API, deploy in a private VPC, run on your own servers or embed detection on a device. Compare the data path and operating requirements.
Where the audio goesAudio crosses into our cloud
Your systems
Your telephonySIP · contact center · application
Encrypted audio
Detectif.ai cloud
Detectif.ai runtimeManaged and scaled by us
Signal response
Your systems
Your policy engineContinue · step up · review
Your infrastructureDetectif.ai model runtime
Hosted API
Start with a managed runtime.
Stream the caller’s audio to Detectif.ai over an encrypted connection. Receive a signal in your application and apply your own decision policy.
- Audio processed in
- Detectif.ai-managed cloud
- Operated by
- Detectif.ai
- Network path
- Encrypted connection to Detectif.ai
- Release control
- Runtime and model updates managed by us
A straightforward starting point for evaluation and cloud-based call flows.
Study reference: 14.4 ms median per clip on an NVIDIA L4 GPU, FP16.Read the deployment study
Use cases.
Add voice checks to onboarding, payments, account recovery and customer service. Explore the call flow for each application.
Choose a call flow
Banks · credit unions · fintech
Voice checks for KYC onboarding.
Add voice authenticity to remote onboarding and identity verification. Inspect the applicant’s speech alongside your document and liveness checks, then enroll a voice reference when the identity process is complete.
- 01Receive the applicant’s speech
- 02Check for synthetic or replayed audio
- 03Complete KYC, then enroll a reference
The checks behind it
Example
Identity document
Government-issued IDPassport or driver’s license
Details maskedIllustrative example
Caller request
“I’d like to open an account.”
Voice authenticitySynthetic score · 0.04
Speaker enrollmentEnroll after your KYC checks
Next stepContinue your onboarding checks
Built for your team
- Banks
- Lenders and credit unions
- Insurers
- Payments and fintech
- Contact centers and BPOs
- Voice AI platforms
Voice model research.
Studies in forensic representations, language transfer and deployment. Explore the methods, results and limitations.
All research
Deepfake detection across devices
One detector, seven deployment configurations. Measure processing capacity and detection quality on GPU, server CPU and Android.
99.65%Of handset test clips processed faster than real time
Questions from risk, fraud and engineering teams.
Straight answers, with the evidence behind each one.
What is voice deepfake detection?
Voice deepfake detection identifies speech that was generated or cloned by AI, or replayed from a recording, instead of spoken live by a person. Detectif.ai’s models score audio for these signs and return an authenticity signal that an application or agent can act on.
How do banks detect cloned voices on phone calls?
By scoring the customer’s side of the call for synthetic speech while it is happening, alongside speaker verification. A good clone can match a voiceprint, so authenticity has to be checked separately before sensitive actions such as payments, credential resets or changes to contact details.
Does it work on real phone-line audio?
Telephone audio is the design target: 8 kHz speech after codecs, packet loss, resampling and background noise, which is how fraud actually arrives. Performance in any deployment depends on channel, language and generator, and is measured on your own traffic during evaluation.
How fast is Detectif.ai?
On the independent Podonos audio deepfake detection benchmark, Detectif.ai reported a mean latency of 24 ms per clip, the lowest of any commercial system that submitted timing. Latency is vendor-reported and depends on hardware and deployment.
How accurate is it?
Podonos scored Detectif.ai at 94.47% accuracy on 4,524 files against labels it holds privately, 9th of 23 systems. 97.3% of synthetic files were caught and 8.4% of genuine files were flagged. Results on your traffic are measured in an evaluation before deployment.
Which languages does it support?
The models are built to be language-agnostic. In Detectif.ai’s cross-lingual study, forensic representations trained on 12 languages reached 91.0% top-1 generator retrieval on 6 languages held out of training. That study measures attribution among known generators; detection performance per language is confirmed during evaluation.
Can it run on-premises or on a device?
Yes. The same models can be hosted, deployed inside your own infrastructure, or run on device. The deployment study measured one detector on an NVIDIA L4 GPU, a server CPU and an Android handset, where 99.65% of clips were processed faster than real time.
Is call audio stored?
Retention is set per deployment. Audio can be analysed without being stored, or the flagged segment can be kept with the model record for review when your policy and regulations allow it.
Test it on your own call traffic.
A 30-minute walkthrough of your highest-risk call flow, then an evaluation on audio from your own lines.