Multilingual Audio DeepFake Detection

Motivation

The overwhelming majority of audio deepfake (DF) detection research is built on English-language datasets such as ASVspoof. Yet synthetic speech technology is language-agnostic — TTS and voice-conversion systems produce convincing fakes in any language. This raises an uncomfortable question: are our best detectors actually polyglots, or are they just good at English?

This project — which forms the core of my Master’s thesis — set out to answer that question systematically.

What We Did

We evaluated a range of state-of-the-art DF detection models on speech spanning 8 languages: English, German, Polish, Ukrainian, Russian, Spanish, French, and Italian. For each language we studied three settings:

1. Zero-shot — model trained only on English, evaluated on a target language.
2. Intra-linguistic adaptation — model fine-tuned and evaluated within the same non-English language.
3. Cross-linguistic adaptation — model fine-tuned on one language and evaluated on another.

We also analysed which acoustic features survive the language shift and which become noise.

Key Findings

• Zero-shot transfer degrades sharply for most languages, with equal-error rates sometimes tripling compared to English performance.
• Intra-linguistic fine-tuning recovers much of the performance, confirming that target-language data is the most important factor.
• Cross-linguistic transfer is inconsistent: languages with shared phonology can partially share knowledge, but there is no reliable free lunch.
• Models that encode language-agnostic artefacts of neural vocoders are more robust; models that latch onto language-specific prosody are brittle.

Takeaways for Practitioners

If you are deploying audio deepfake detection in a non-English environment, you should not assume English-trained baselines will protect you. Collecting even a small amount of target-language data for adaptation can make a decisive difference.

Related Publication

The research is described in detail in Are Audio DeepFake Detection Models Polyglots? (arXiv 2024).