Skip to main content

Understanding RealCall Results

Learn how to read RealCall detection results, including call conclusions, score ranges, audio segments, inconclusive reasons, and model signals.

Written by Wen Huang

When a call finishes processing, RealCall returns a conclusion for the overall call, plus a chronological breakdown of the individual audio segments that were analyzed.

This article explains how to read RealCall results, including the overall conclusion, segment-level findings, inconclusive reasons, model tags, and related settings.

Overall call result

RealCall’s overall call result is shown as the Conclusion. It reflects the likelihood that the call contained manipulated or AI-generated audio. You can find the Conclusion field in the RealCall menu and on each call’s detail page.
​
RealCall can return four conclusion values:

  • Authentic (0 - 55) — the call appears to contain a real human voice.

  • Suspicious (55 - 65) — the call shows signs of manipulation, but not enough to meet the artificial/manipulated threshold.

  • Manipulated / Artificial (65 - 100) — the call appears to contain AI-generated or manipulated audio.

  • Inconclusive — analysis ran, but RealCall could not reach a confident determination.

RealCall scores each analyzed segment on a 0–100 scale, where higher scores indicate a higher likelihood that the speech is artificial. Inconclusive results do not map to a score range; see the Inconclusive reasons section below for more detail.

Individual call results

Opening a call displays a dedicated results page. The page includes:

  • The call's direction, date and time, call type, call ID, time to verdict, and detected language

  • The overall conclusion panel

  • Segment-level results, playback, and download controls

Call type shows whether the result is from an Actual call processed live by RealCall or a Test call created from an uploaded audio file.
​
The previous/next controls let users move between results without returning to the list. Filters, search, and sort context carry over while navigating.
​

Audio segments

RealCall processes the call audio in short chunks, then combines chunks with enough usable speech into 3-second audio segments for analysis. Chunks without enough speech, such as silence, are not used to form analyzed segments.

RealCall uses segment-level results to determine the overall call result. A call-level Conclusion is drawn when RealCall finds two segments with the same result. Because skipped silence or other skipped audio may occur between analyzed segments, the time to verdict can be longer than the total duration of the segments shown.

The segment list helps explain which parts of the call contributed to the final Conclusion and which models agreed.

Segment states

Each segment can show one of these states:

  • Authentic — the segment appears to contain a real human voice.

  • Manipulated — the segment appears to contain AI-generated or manipulated audio.

  • Skipped — the segment was not analyzed because it did not contain enough usable evidence. Skipped segments are treated as inconclusive, and RealCall may show a reason when one is available. See Inconclusive reasons for more detail.

Segments with an Authentic or Manipulated result also show which RealCall detection models contributed to the result:

  • Everest (called Advanced in RealScan): Uses general-purpose embeddings that discriminate synthetic from real human speech.

  • SLIM (called Generalizable in RealScan): Uses mismatch in style and linguistic patterns that are otherwise unique to real human speech.

  • Foundational: Uses embeddings from Reality Defender’s proprietary encoder, trained to learn deepfake artifacts directly rather than adapted from general speech models.

Inconclusive reasons

When RealCall cannot use a segment or cannot reach a confident call-level result, it provides a plain-language reason. The same reason is returned by the API as a reason code.

The API reason codes are:

Code

Meaning

cross_talk

Multiple speakers or overlapping audio made the segment difficult to analyze.

dialtone_or_music

The audio primarily contained dial tone, hold music, or similar non-speech audio.

insufficient_evidence

The available signal was not strong enough for a confident determination.

insufficient_speech

Not enough usable speech to reach a conclusion.

low_quality_audio

Audio quality was too poor for a confident result.

mixed

The audio contained mixed signals or mixed conditions that prevented a clear result.

other

RealCall could not classify the reason more specifically.

recognized_system_voice

The audio matched a recognized system or automated voice.

silence

Mostly silence or not enough speech to analyze.

unsupported_language

The spoken language is not supported for this analysis.

Upload test audio

Admins and Contributors can upload a test audio file from the RealCall menu for analysis. When upload succeeds, the result appears in the calls list as a Test call.
​

The upload flow includes:

  • Drag-and-drop or file browse

  • Supports MP3, WAV, and AIFF audio files up to 20 MB

  • A Simulate a phone call toggle that appears after a file is selected

  • The account's RealCall test phone number

The test phone number lets users place a live test call without uploading a file. The Simulate a phone call option processes an uploaded file through a phone-call-like workflow, so users can test how the audio behaves under telephony conditions.

Need help?

  • Chat: Ask the bot in the bottom right corner first. If needed, escalate to support from the same chat.

  • Email: Send details about your question or issue to support@realitydefender.com.

Related Articles
Did this answer your question?