DiffWave Diffusion Vocoder Quiz

Answers and explanations appear as you take the quiz.

1. How does DiffWave generate audio at inference time?
   • By iteratively denoising random noise into a waveform
   • By sampling one waveform value at a time autoregressively
   • By concatenating recorded speech units
   • By running a single GAN forward pass
2. What does the DiffWave network actually learn to predict during training?
   • The clean waveform directly in one shot
   • The noise added at a given diffusion step
   • The text transcript
   • The speaker identity
3. What is DiffWave's main practical drawback compared to GAN vocoders?
   • It cannot condition on spectrograms
   • It only works at low sample rates
   • It can require many sampling steps, making it slower
   • It needs adversarial training to be stable
4. What advantage does DiffWave have over GAN vocoders in training?
   • It needs no neural network
   • It requires no data
   • It produces audio instantly in one step
   • It avoids unstable adversarial training
5. What architecture does DiffWave's noise-prediction network resemble?
   • A WaveNet-style dilated convolutional network
   • A recurrent LSTM language model
   • A vision transformer
   • A simple linear regression
6. Roughly how few sampling steps can well-tuned DiffWave schedules use while keeping good quality?
   • Exactly 1000 steps always
   • About six steps
   • Around 500 steps
   • Just one step with no tuning
7. What acoustic representation can condition DiffWave when it is used as a vocoder?
   • A list of subtitle line numbers
   • A CT scan of an ancient scroll
   • A mel-spectrogram
   • A source-code dependency graph
8. What loss does DiffWave use to train its noise-prediction network?
   • A word-error-rate objective
   • A discriminator feature-matching loss only
   • A cosine-similarity loss between image embeddings
   • A weighted L2 loss on the predicted noise