Noise2Noise Speech Enhancement Quiz

Answers and explanations appear as you take the quiz.

1. What is the surprising core claim of Noise2Noise?
   • You can denoise using only corrupted examples, no clean references needed
   • You must always have clean ground-truth audio
   • Noise cannot be removed by neural networks
   • Only speech, never images, can be denoised this way
2. Why can't the network simply memorize the noise in the target clip?
   • The noise is too quiet to matter
   • The noise is random and independent, so it is unpredictable
   • The network is too small
   • The noise is always the same value
3. For squared-error training, what conditional statistic minimizes expected loss?
   • The largest target value
   • The unconditional variance
   • The conditional mean of the target given the input
   • A randomly selected target frame
4. What must be true about the noise added to the target for the trick to work?
   • It must be louder than the speech
   • It must be the same noise as the input
   • It must come from a recording studio
   • It must be zero-mean and independent of the input's noise
5. For absolute-error training, which conditional statistic is minimized?
   • The conditional median of the target given the input
   • The variance
   • The maximum
   • The sum of all pixels
6. Which methods extended Noise2Noise toward learning from single noisy samples?
   • Word2Vec and BERT
   • Noise2Void and Noise2Self
   • Griffin-Lim and SoundStream
   • ResNet and VGG
7. Which non-photographic reconstruction task did the Noise2Noise paper demonstrate?
   • Satellite orbit prediction
   • Particle-track reconstruction
   • Reconstruction of undersampled MRI scans
   • Speech translation
8. Which condition can undermine Noise2Noise’s clean-target argument?
   • The model uses a convolutional layer
   • The input and target have different file names
   • The clean signal contains edges
   • The target corruption is predictably biased given the input corruption