Product
Introducing Aletheia: Audio Deepfake Detection, Now on the Detectif.ai Platform
Aletheia, our audio deepfake detection model, is now available to everyone on the Detectif.ai platform, with pay-as-you-go pricing and 2,000 free credits for new accounts.
Aletheia detects synthetic and cloned speech in live calls and recordings. It returns a synthetic score for the audio and marks the segments that need review.
In our internal evaluation:
- Clean English speech: 98.3% accuracy. It catches 98.3% of AI voices and flags 1.7% of real ones.
- Noisy, compressed, echoey, VoIP and phone audio: 97% accuracy in English.
- Real phone calls with AI voice agents: it caught 152 of 152 in one set and 99.4% of 1,015 in a second set.
On the independent Podonos audio deepfake detection benchmark, Aletheia reported a mean latency of 24 ms per clip, the lowest of any commercial system with published timing. Podonos scored it at 94.47% accuracy across all 4,524 files, with 97.3% of deepfakes detected.
Aletheia is available in the platform dashboard and through the API for every account, and new accounts get 2,000 free credits. It is also on X: tag @DetectIfAI under a post with audio or video and it replies with what it found. Pricing is at the end of this post. For enterprise volumes and private deployment, talk to us directly.
Why Now?
Voice cloning and text-to-speech now sound close to human. Systems such as ElevenLabs v3 and v4, Hume, Speechify and Kokoro produce speech that is hard to tell from a real voice, and a familiar voice can be cloned from a few seconds of audio.
AI voice agents already make phone calls at scale. Cloned voices are used to take over accounts through the contact center, to authorise payments by impersonating an executive, a relative or a customer, and to pass remote onboarding with synthetic speech.
These attacks arrive over the phone, where a good clone is hard to hear and a decision is made in seconds. A deepfake score is useful only if it arrives while that decision is still open.
Detection Capabilities
Aletheia scores speech for two kinds of manipulation:
- Synthetic speech: audio generated by text-to-speech systems, and
- Voice cloning: a real person’s voice reproduced from a recording.
It runs on live calls and on recordings. On a live call it scores the caller’s side of the conversation as the call happens. On a recording it returns the score with the flagged segments, so a reviewer can go straight to the part that matters.
Telephone audio is the design target: 8 kHz speech after codecs, packet loss, resampling and background noise. The models are built to be language-agnostic.
Performance in any deployment depends on the channel, the language and the generator, so we measure it on your own audio during an evaluation.
Built for Live Calls
Aletheia’s models are small. Each check needs little compute, which keeps latency low and keeps the cost of running them low, on GPU or CPU.
Our current model runs at 24 ms on GPU and 84 ms on CPU.
The same models run in the cloud, in a private VPC, on-premises or on-device, so audio can stay inside your own infrastructure.
Evaluation and Testing
Clean and Realistic Audio
On clean English speech, Aletheia reaches 98.3% accuracy: it catches 98.3% of AI voices and flags 1.7% of real ones. On clean Hindi speech it reaches 98.5%.
Calls rarely arrive clean. Across noisy, compressed, echoey, VoIP and phone audio, accuracy is 97% in English and 96.4% in Hindi.
Real Calls with AI Voice Agents
We tested Aletheia on real phone calls placed by AI voice agents. It caught 152 of 152 calls in one set and 99.4% of 1,015 in a second set.
Independent Benchmark
Podonos runs a public benchmark for audio deepfake detection. It hosts a fixed evaluation set of 4,524 files, holds the labels privately and computes every score itself, so no vendor scores its own row. The September 2026 board covers 23 systems.
- Detectif.ai24 ms
- Pella Research57 ms
- Pindrop282 ms
- Resemble DETECT-World399 ms
Aletheia reported 24 ms per clip, or 119× faster than real time: 2.3× the processing speed of the next-fastest commercial system. Podonos scored the same entry at 94.47% accuracy. The full write-up is in our Podonos post, and the board is in the Podonos write-up.
Every File Scored
Aletheia returned a decision on all 4,524 files and declined none. Some commercial systems declined part of the set, mostly short clips, so their accuracy covers a smaller sample.
Faster Than Real Time
Across Languages
In our 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, not detection accuracy.
How to Try It
Create an account on the Detectif.ai platform. Every new account gets 2,000 free credits to start.
On X, tag @DetectIfAI under a post with audio or video. The bot checks the audio and replies with one of three results: no AI detected, AI detected, or partly AI and inconclusive.
Upload audio in the dashboard, or follow the Quickstart to go from an API key to a first deepfake check. The API reference covers recordings and live audio.
Pricing
Platform prices are pay as you go, in US dollars. Enterprise terms are agreed directly with our team.
Pay as you go
For developers and teams. Sign up, add a payment method and call the API or use the dashboard.
| Service | Unit | Price |
|---|---|---|
| Deepfake detectionRecordings and live audio | Per second analysed | $0.0002$0.012 per minute · $0.72 per hour |
| Speaker enrolment | Per enrolment | $0.02 |
| Speaker verification | Per call | $0.05 |
| Call scam analysis | Per call, up to 15 minutes | $0.10 |
| BundleScam analysis, speaker verification and deepfake detection | Per call, up to 15 minutes | $0.15 |
2,000 free credits. Every new account gets 2,000 credits to start.
Start on the platformTalk to us
For banks, contact centers and platforms with high call volumes or their own infrastructure.
- Volume pricing on a yearly commitment
- Cloud, private VPC, on-premises or on-device deployment
- A pilot sized to your traffic