3 ms·
From the abstract: Given any audio waveform, we can produce another that is over 99.9% similar, but transcribes as any phrase we choose (at a rate of up to 50 c
by pulisse 9y ago
From the abstract: Given any audio waveform, we can produce another that is over 99.9% similar, but transcribes as any phrase we choose (at a rate of up to 50 characters per second).
- yorwba 9y agoThat sentence is quite easy to misunderstand. The 50 characters per second are the capacity of the attacked model (it chops the audio in frames of 1/50th second), and what they want to highlight is that they can cram the maximum amount of text into the adversarial example. The actual generation takes an hour on an nVidia 1080Ti, but they can parallelize the process to compute multiple examples on the same GPU, giving an amortized cost of a few minutes.