Detecting a fake voice is already a major challenge. But what if we could go further — and actually understand how it was created and even who might be behind it?
From tasting wine to tracing voices
Imagine you’re tasting a glass of wine.
Most people can say whether they like it or not.
That’s detection: this sounds authentic / this sounds fake.
But a trained sommelier can go much further. With a single sip, they can tell you the grape variety, the region, sometimes even the process used to produce it.
That’s the leap from detection to source tracing.
What source tracing really means
Until now, most audio deepfake systems could only answer a binary question: is this real or fake?
Important, yes — but limited.
Source tracing goes deeper. It analyzes the building blocks of a synthetic voice:
- the acoustic model used to shape pronunciation and rhythm,
- the vocoder that generated the raw waveform,
- and the overall system type that combined them.
By identifying these elements, we don’t just know that something is fake. We understand how it was made — and in some cases, we can even link it back to the group, toolset, or actor likely responsible for the attack.
Why it matters
- Greater visibility – Understanding the method provides a clearer map of the threat landscape.
- Attribution of attacks – Certain combinations of models or tools leave behind unique “fingerprints,” making it possible to associate a fake with a specific attacker or family of attacks.
- Proactive defense – Knowing the origin method helps anticipate the next moves of the same adversary.
- Trust and credibility – In sensitive industries, being able to explain not only that an audio is fake but also who could be behind it provides unmatched assurance.
Conclusion
Deepfake detection was the first milestone.
Source tracing is the next chapter — one that lets us see inside the fake, understand its components, and in some cases, even point to the actor behind the attack.
Because just like with wine: anyone can say whether it tastes good or not.
But only with the right expertise can you tell the grape, region, process — and even the vineyard behind it.
Source :
Audio Deepfake Source Tracing using Multi-Attribute Open-Set Identification and Verification
Pierre Falez¹ Tony Marteau¹, Damien Lolive², Arnaud Delhay³
¹ Whispeak, France
² Univ Bretagne Sud, CNRS, IRISA, France
³ Univ Rennes, CNRS, IRISA, France
