I am a software engineer currently developing a somewhat innovative audio watermarking algorithm for a bootstrapped company named Parseval that I am launching with a few friends, and we’d like to gather some signal on whether or not people can hear the watermark :)
Our goal with this watermark, is to help identifying masters when they're reused, including in AI training data, to protect artist's original work.
The test we designed is using the 2AFC with labelled reference method because with a known reference it is basically more sensitive than regular and more common ABX (see Hautus et al., 2011) test. Basically, this is like ABX except that the reference is labelled and you have to choose between two unlabelled versions, which removes the memory burden of X.
Note: it is highly recommended to perform this test on headphones, you'll probably hear things more clearly and sharply. There are a few checks at the beginning of the test so that we can make sure everything is working fine on your side :)
The test is simple:
- It works directly in your browser, requires absolutely no signup or anything and takes about 10-15 minutes. We used LLMs to make the test available in 10 different languages, sorry if the translations are not really accurate, and feel free to report this to me directly here.
- Each round during the test, you’ll have one original excerpt, and then two different versions, one being the original and the other one being watermarked. Your goal is to pick the watermarked one.
- There are some special rounds, typically rounds where the watermark is boosted (from +6dB to +36dB), and control rounds with added white noise.
- All the audio comes from great artists and CC-licensed music, and they are all credited, so if you like a song, please feel free to check out the artists pages :)
At the end of our study, we’ll make sure to release the method for the analysis as well as the aggregated results and of course this, whatever the result. Those results will help us understand if trained people can detect it and might help us improve our algorithm.
I would be super happy to answer any questions, about the watermark, our method or anything you'd like to ask :)
gabyfle•51m ago
I am a software engineer currently developing a somewhat innovative audio watermarking algorithm for a bootstrapped company named Parseval that I am launching with a few friends, and we’d like to gather some signal on whether or not people can hear the watermark :)
Our goal with this watermark, is to help identifying masters when they're reused, including in AI training data, to protect artist's original work.
The test we designed is using the 2AFC with labelled reference method because with a known reference it is basically more sensitive than regular and more common ABX (see Hautus et al., 2011) test. Basically, this is like ABX except that the reference is labelled and you have to choose between two unlabelled versions, which removes the memory burden of X.
Note: it is highly recommended to perform this test on headphones, you'll probably hear things more clearly and sharply. There are a few checks at the beginning of the test so that we can make sure everything is working fine on your side :)
The test is simple: - It works directly in your browser, requires absolutely no signup or anything and takes about 10-15 minutes. We used LLMs to make the test available in 10 different languages, sorry if the translations are not really accurate, and feel free to report this to me directly here. - Each round during the test, you’ll have one original excerpt, and then two different versions, one being the original and the other one being watermarked. Your goal is to pick the watermarked one. - There are some special rounds, typically rounds where the watermark is boosted (from +6dB to +36dB), and control rounds with added white noise. - All the audio comes from great artists and CC-licensed music, and they are all credited, so if you like a song, please feel free to check out the artists pages :)
At the end of our study, we’ll make sure to release the method for the analysis as well as the aggregated results and of course this, whatever the result. Those results will help us understand if trained people can detect it and might help us improve our algorithm.
I would be super happy to answer any questions, about the watermark, our method or anything you'd like to ask :)