HiFi-GAN and GAN Vocoders Quiz

Answers and explanations appear as you take the quiz.

1. What is the primary job of a vocoder like HiFi-GAN in a text-to-speech pipeline?
   • Turning a mel-spectrogram into a raw audio waveform
   • Converting raw text into phonemes
   • Aligning text to durations
   • Translating speech from one language to another
2. Why is HiFi-GAN much faster than the earlier WaveNet vocoder?
   • It uses a smaller vocabulary
   • It is feed-forward and produces the whole waveform in a single pass instead of autoregressively
   • It generates samples one at a time more efficiently
   • It skips the spectrogram entirely
3. What does HiFi-GAN's Multi-Period Discriminator specifically help capture?
   • The semantic meaning of words
   • Background noise levels
   • Periodic patterns related to pitch in the signal
   • The speaker's emotional tone label
4. GAN vocoders are trained 'adversarially.' What does that mean here?
   • The model competes against human narrators
   • Two languages are mixed during training
   • The model is rewarded only for speed, not quality
   • A generator tries to fool one or more discriminators that judge whether audio is real
5. Which later vocoder extends HiFi-GAN to generalize better to unseen voices, singing, and instruments?
   • BigVGAN
   • Tacotron
   • WaveNet
   • BERT
6. What auxiliary loss helps keep HiFi-GAN's training stable and faithful to the input?
   • A translation loss
   • A mel-spectrogram reconstruction loss
   • A token-classification loss
   • A reinforcement reward signal
7. What does HiFi-GAN’s Multi-Scale Discriminator examine?
   • Only the text transcript
   • The model’s optimizer state
   • Waveforms at several downsampled resolutions
   • One mel bin at a time
8. Why include a mel-spectrogram reconstruction loss alongside adversarial loss?
   • To eliminate the generator
   • To make the model autoregressive
   • To remove the conditioning mel-spectrogram
   • To anchor spectral content while discriminators encourage realistic waveform texture