MelGAN Generative Vocoder Quiz

Answers and explanations appear as you take the quiz.

1. What input does MelGAN convert into a raw audio waveform?
   • A mel-spectrogram
   • Raw text characters
   • A MIDI file
   • A phoneme lattice
2. Why is MelGAN much faster than WaveNet at inference?
   • It uses a smaller sample rate
   • It is non-autoregressive and generates audio in one parallel pass
   • It runs only on TPUs
   • It skips the high-frequency bands
3. What distinctive discriminator design did MelGAN introduce?
   • A single global discriminator
   • A transformer-based discriminator
   • A multi-scale discriminator operating on audio at several resolutions
   • No discriminator at all
4. Which loss helps stabilize MelGAN's adversarial training?
   • Cross-entropy on text tokens
   • Mean opinion score loss
   • Triplet margin loss
   • Feature-matching loss between discriminator activations
5. How does MelGAN's generator increase its receptive field?
   • By using dilated convolutions in residual blocks
   • By adding recurrent LSTM layers
   • By using attention over the whole clip
   • By stacking pooling layers
6. What subjective evaluation measure did the MelGAN paper report?
   • Word error rate (WER)
   • Mean Opinion Score (MOS)
   • BLEU score
   • Peak signal-to-noise ratio (PSNR) only
7. What do MelGAN’s transposed convolutions do in the generator?
   • Convert the waveform back into text
   • Choose the phonemes in the input sentence
   • Upsample the mel-spectrogram representation toward the audio waveform rate
   • Reduce the number of audio samples to zero
8. Which audio domains did the MelGAN paper use to illustrate the generality of its techniques?
   • Machine translation, image captions, and video subtitling
   • Only speech recognition and speaker identification
   • Text summarization, code generation, and image classification
   • Speech synthesis, music-domain translation, and unconditional music synthesis