Who can use this feature?
Plan: Call Center AI
Managed from: Admin Portal
User type: Admin, User
Deepfake Detection analyzes customer speech in real time during a call and identifies whether the agent is talking to a real human or to an AI-generated (synthetic) voice. When a synthetic voice is identified, Krisp flags the call, collects detection analytics, and can alert the agent so they can follow your team's handling procedures.
Info
Deepfake Detection is a paid add-on that can be connected to any Call Center AI plan. To enable it for your team, contact the Krisp Sales team.
How Deepfake Detection works
Detection runs automatically from the moment a call begins. Krisp listens to the customer side of the conversation and processes speech in 4-second chunks, sending each chunk to the cloud for analysis as soon as it is collected. Analysis runs entirely in the cloud, so there is no noticeable impact on the agent's CPU.
Detection runs before any Krisp voice processing, so features like Voice Translation and Accent Conversion do not cause false positives. Detection stops after a successful result and restarts when a new call is received. Agents can also trigger the analysis manually at any point during a call if something seems off.
Info
Krisp Cloud has access to customer audio only in transit and never stores it. Audio chunks are sent for analysis with real-time encryption.
Setting up Deepfake Detection
Once the add-on is enabled for your account, an Admin configures it in the Admin Portal.
- Go to Team settings >>> Products >>> Voice security.
- Turn on Deepfake detection
- Enable any additional settings your team needs (Detection delay, Real-time alerting, Play whisper, Allow running in debug mode).
Detection becomes active on all calls right away. No calibration or training is required.
Deepfake Detection settings
| Setting | What it does |
|---|---|
| Deepfake detection | Master switch. Analyzes all calls for potential deepfake risk and collects analytics. |
| Detection delay | Delays detection after the call starts to avoid false positives from opening whispers or prompts. Choose 5, 10, 15, or 20 seconds; the default is 10 seconds. |
| Real-time alerting | Warns the agent during an active call when a deepfake risk is detected. If this is off, nothing is shown to the agent during the call. |
| Play whisper when synthesized voice is detected | Plays a short whisper to the agent when a synthetic voice is detected. Requires Real time alerting to be on. |
| Allow running Deepfake Detection in debug mode | Lets Krisp collect and analyze audio chunks so detection issues can be investigated through problem reports. Turn this on when troubleshooting a detection issue with the Krisp team. |
Verdicts and agent alerts
After analyzing a chunk of customer speech, Krisp returns one of the following verdicts. Agents see a live alert in the Krisp widget only when real-time alerting is on.
| Verdict | What it means | What the agent sees |
|---|---|---|
| Confirmed human voice | Strong indicator of a human voice. | A green confirmation: "You are talking to a human." Shown after a manual check. |
| Potential deepfake risk | Something abnormal was detected, but confidence is not high enough for a definitive result. Treat with caution. | A yellow warning prompting the agent to pause sensitive actions and follow company instructions. |
| High deepfake risk | Strong indicator of a synthetic or AI-generated voice. | A red warning prompting the agent to pause sensitive actions and follow company instructions. |
| Detection failed, retry | The analysis could not produce a result, usually due to audio quality or network issues. | A message that customer voice detection failed, with a Retry option. |
Important
Krisp flags the risk but does not decide how to handle the call. When a deepfake risk is detected, agents should follow their company's internal handling procedures.
Reviewing detections in the Admin Portal
If your team also uses Speech Analytics, when Deepfake detection is on, every call where a deepfake was detected is marked in the Call history list, so Admins get a quick view of flagged activity across all calls. From the call details view, an Admin can review the verdict and mark a call as "Not a deepfake" if the flag is believed to be incorrect.
Analytics dashboard
The Deepfake Detection analytics dashboard is available to Admins and Analysts in the Admin Portal. It gives a high-level view of synthetic voice activity across all calls:
- Total detections: the number of calls flagged as deepfake risk in the selected period
- Detection ratio: the percentage of flagged calls out of total calls handled
- Detection trend: a chart of flagged call volume over time, filterable up to a 90-day window.
- Detections list: breaks activity down by individual call. Each row shows the user's email or device, username, the date and time of the call, the detection type (Suspected or High-risk), and the confidence score for that detection.
FAQ
It is a paid add-on that can be connected to any Call Center AI plan. Contact the Krisp Sales team to enable it.
No. Deepfake Detection is built into the application and becomes available in the Admin Portal based on your plan.
In the cloud. Krisp collects audio chunks and sends them for analysis with real-time encryption. The audio is accessed only in transit and is never stored.
The feature is language-agnostic and works across all languages with no configuration required.
Detection runs in the cloud, so there is no noticeable impact on CPU performance.
Detection starts automatically each time a call start is detected. In environments using a persistent connection, Krisp cannot identify the call start unless the Krisp Bridge is enabled or the agent manually ends each call.
Handling varies by company. Agents should follow their organization's internal procedures for suspected fraud.