Description
Automatic Speech Recognition (ASR) systems are so abundant and plentiful nowadays, that it is difficult not to
deal with one a few times every day. Most state-of-the-art ASR systems are vulnerable to adversarial examples
that can fool the system into transcribing whatever the attacker wants. Previous work examined the vulnerability
of English ASR systems against various types of adversarial attacks, but very few others have compared the susceptibility
of ASR systems of a different languages to adversarial attacks. In this work, we inspect the susceptibility of multiple
French ASR models to numerous types of adversarial attacks and we compare our results with previous literature.
Our evaluation indicates that there is a statistically significant difference between French, English and German ASR
systems. Moreover, we show that the Word Error Rate (WER) of a model has significant effect on its attackabilty.
|
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching