UnivNet Multi-Resolution Vocoder Quiz

Answers and explanations appear as you take the quiz.

1. What is the signature feature of UnivNet's discriminator?
   • It computes multiple STFT spectrograms at different resolutions and discriminates on them
   • It uses a single waveform discriminator only
   • It judges audio using a language model
   • It ignores high frequencies
2. What problem in earlier GAN vocoders does UnivNet specifically target?
   • Slow autoregressive sampling
   • Muffled or artifact-laden high frequencies
   • Inability to run on GPU
   • Lack of any discriminator
3. What are location-variable convolutions (LVC) in UnivNet?
   • Convolutions applied only at the start of the signal
   • Fixed kernels shared across all layers
   • Convolutions whose kernels are predicted dynamically from input features
   • Recurrent connections between frames
4. What adaptation potential did the UnivNet paper report for new speakers?
   • Guaranteed identical quality across every recording condition
   • No need for conditioning features or evaluation
   • Instant adaptation to every language and microphone
   • Fast adaptation without training a new vocoder from scratch
5. What synthesis capability did the UnivNet paper report?
   • High-fidelity waveform synthesis in real time
   • Only offline synthesis after a day-long wait
   • One waveform sample per hour
   • A text transcript without audio output
6. What does UnivNet's generator condition on to produce a waveform?
   • Raw text
   • A mel-spectrogram
   • A spectrogram of noise only
   • A phoneme duration table
7. Which adversarial loss did UnivNet use to train its generator and discriminators?
   • A word-level cross-entropy loss
   • A policy-gradient objective
   • A least-squares GAN objective
   • A contrastive image-text loss
8. What evaluation result did the UnivNet paper report across seen and unseen speakers?
   • It outperformed every vocoder on every language and dataset
   • It proved no speaker adaptation is ever needed
   • It achieved the same result regardless of input features
   • It achieved the best objective and subjective results among the compared models in its evaluation