Parallel WaveGAN Vocoder Quiz

Answers and explanations appear as you take the quiz.

1. What is the job of a vocoder like Parallel WaveGAN?
   • Turn a mel-spectrogram into a raw audio waveform
   • Convert text into phonemes
   • Remove background noise from recordings
   • Translate speech between languages
2. How does Parallel WaveGAN generate audio samples compared with original WaveNet?
   • One sample at a time, sequentially
   • All samples in parallel, non-autoregressively
   • By recording a microphone
   • By concatenating stored clips
3. Which loss is central to Parallel WaveGAN besides the adversarial loss?
   • Cross-entropy on text tokens
   • Triplet ranking loss
   • Multi-resolution STFT loss
   • Reinforcement reward
4. What architecture does the Parallel WaveGAN generator use?
   • A recurrent LSTM
   • A vision transformer
   • A decision tree
   • A dilated-convolution WaveNet-style network
5. Why is Parallel WaveGAN attractive for deployment?
   • It is compact and runs faster than real time
   • It is a very large multi-billion-parameter model
   • It requires no spectrogram input
   • It only works on supercomputers
6. In a GAN, what role does the discriminator play during training?
   • It generates the spectrogram
   • It judges whether audio looks real or fake
   • It records the dataset
   • It stores model weights
7. In Parallel WaveGAN, what does the noise input contribute alongside mel-spectrogram conditioning?
   • The speaker transcript after synthesis
   • A replacement for the discriminator
   • Stochastic input for waveform generation while the mel features condition the acoustic content
   • A fixed random waveform copied into every output
8. What does the adversarial loss encourage the Parallel WaveGAN generator to produce?
   • Text transcripts with perfect spelling
   • Audio samples with no speaker conditioning
   • A smaller mel-spectrogram vocabulary
   • Waveforms that the discriminator judges as resembling real speech