WaveGlow Flow-Based Vocoder Quiz

Answers and explanations appear as you take the quiz.

1. WaveGlow combines architectural ideas from which two models?
   • Glow and WaveNet
   • BERT and GPT
   • ResNet and VGG
   • Tacotron and CLIP
2. What kind of model is WaveGlow?
   • A generative adversarial network
   • A normalizing flow
   • A recurrent autoregressive model
   • A reinforcement learning agent
3. How does WaveGlow generate a waveform at inference time?
   • It predicts samples one by one
   • It searches a database of clips
   • It samples Gaussian noise and runs the flow in reverse
   • It uses a discriminator
4. What objective does WaveGlow optimize during training?
   • An adversarial min-max game
   • Cross-entropy over text labels
   • A contrastive loss
   • Exact log-likelihood of the audio data
5. Which two components make up each invertible step in WaveGlow?
   • Affine coupling layer and invertible 1x1 convolution
   • Attention and feed-forward block
   • Pooling and softmax
   • Embedding and dropout
6. What training simplification does WaveGlow offer over earlier parallel WaveNet methods?
   • It needs no spectrogram
   • It avoids teacher-student distillation by using a single network and loss
   • It removes the need for any GPU
   • It does not need audio data
7. What does WaveGlow's squeeze operation do before its coupling layers process the signal?
   • Converts the waveform into written phonemes
   • Removes the mel-spectrogram conditioning
   • Groups neighboring audio samples into vectors for the flow transformations
   • Randomly deletes audio segments
8. Which acoustic model is named in the guide as a practical system that can pair with WaveGlow?
   • BERT
   • A* search
   • YOLOv8
   • Tacotron 2