Audio Manipulation Detection

Phonexia has developed Audio Manipulation Detection technology designed to identify whether an audio recording has been deceptively manipulated — for example, by copying and pasting segments into different parts of the recording.

Perpetrators may meticulously splice together audio snippets from various sources to construct misleading narratives or to falsely incriminate individuals.

Possible use cases

• Forensic audio analysis: Identifying tampered evidence in legal investigations, ensuring the integrity of audio submitted in court.
• Call center fraud prevention: Monitoring for tampered customer recordings that could indicate attempts at impersonation or fraudulent behavior.

Scoring

Score values range from negative infinity to positive infinity. The score is a Log-Likelihood Ratio (LLR), which measures the strength of the evidence supporting either hypothesis:

• Values closer to negative infinity suggest the evidence is more likely under Hypothesis 0 ("not manipulated").
• Values closer to positive infinity suggest the evidence is more likely under Hypothesis 1 ("manipulated").

The output score is calibrated such that 0 corresponds to the Equal Error Rate (EER) point on our evaluation datasets. The EER is the point at which the false acceptance rate and false rejection rate are equal, providing a balanced trade-off between the two.

Depending on your specific use case and the characteristics of your data, you may need to adjust the decision threshold to achieve the desired balance between false positives and false negatives.