Learn the difference between TTS and Voice Conversion (VC), how our audio models look for signs of AI-generated speech, how their scores are combined to determine a result, and what factors, such as noise, speaker count, and clip length, can affect detection.
Supported languages
Our audio and voice detection models currently support the following languages:
English
French
German
Italian
Japanese
Korean
Mandarin Chinese
Portuguese
Russian
Spanish
For best results, submit audio in one supported language at a time.
How RD Detects Audio Deepfakes
Reality Defender analyzes two primary types of AI-generated voice content:
Text-to-Speech (TTS):
Converts written text into speech using a target speaker’s voice. Modern TTS systems can convincingly clone a voice from just a few seconds of recorded speech.Voice Conversion (VC):
Takes speech from a source speaker and alters it to sound like a target speaker — essentially impersonation. VC models also need only a few seconds of the target’s voice for convincing synthesis.
In some rare cases, TTS and VC can be combined, though these hybrid approaches are typically less efficient.
Our Detection Methodology
Reality Defender analyzes audio in short segments using multiple specialized models. Rather than relying on spectrogram images alone, the models look for different signs of AI-generated or manipulated speech, such as unusual speech patterns and acoustic artifacts.
Each model produces a score for an audio segment.
An ensemble combines the model scores into an overall likelihood score.
The score indicates the likelihood that audio is AI-generated or manipulated; it isn’t a guarantee or, by itself, an explanation of which signals led to the result.
How Final Audio Classification Is Determined
Reality Defender’s audio detection pipeline uses an ensemble system of three specialized neural models.
Step 1: Segmentation
Audio files are divided into segments for efficient parallel processing.
Each segment is independently evaluated by each model.
Step 2: Independent Scoring
Each model outputs its own fake/real probability score for each segment.
Step 3: Ensemble Decision Rule
The model outputs are combined by an ensemble model that makes the final classification.
To classify an entire audio file as fake, at least two consecutive segments must be marked as fake.
This rule helps minimize false positives, ensuring that genuine audio is rarely misclassified.
Note: The current system is not optimized to detect partial fakes (where only part of an audio clip is manipulated). The classifier determines authenticity for the entire file based on overall signal consistency.
Related FAQs
Question | Answer |
What types of audio deepfakes are detected? | TTS (text-to-speech) and VC (voice conversion). |
How confident are detections? | Each model produces a 1–99% confidence score; the ensemble aggregates them for final classification. |
Can you detect overlapping speakers? | Not yet — the current system focuses on single-speaker scenarios. |
Can you tell if it’s TTS vs voice-conversion? | No, differentiation between TTS and VC isn’t yet available. |
Do you create voice prints or verify identity? | No. Reality Defender only detects whether content is synthetic — not who produced it. |