Responsible use

asr-attacks is a robustness-evaluation library. Use it only on models and recordings you are authorized to test: your own checkpoints, licensed benchmarks, or systems whose owners have asked for a white-box evaluation.

Do not use it to interfere with production speech systems, voice authentication, or other people's devices.

Attacks here are white-box: they need differentiable CTC logits. They are not a recipe for attacking a closed API.

If you publish numbers, cite this package (CITATION.cff), the attack papers you actually implemented, and IBM ART when you use the imperceptible stage. Read algorithms for how the audio CW attack maps to Carlini & Wagner (2018), not the 2017 image C&W paper.

Security issues in the library itself: see SECURITY.md.