Neural Vocoders Quiz

Answers and explanations appear as you take the quiz.

1. What is the primary job of a neural vocoder?
   • To turn an acoustic representation like a mel-spectrogram into an audible waveform
   • To convert text directly into meaning
   • To translate speech between languages
   • To remove background noise from recordings
2. Which 2016 model was the breakthrough that made neural vocoders sound natural?
   • HiFi-GAN
   • WaveNet
   • Tacotron
   • MelGAN
3. Why was the original WaveNet slow to generate audio?
   • It required a separate GPU for each word
   • It re-trained itself on every input
   • It predicted audio one sample at a time autoregressively
   • It rendered video alongside the audio
4. What information does a mel-spectrogram notably discard, making the vocoder's task harder?
   • The loudness of the audio
   • The language being spoken
   • The sample rate
   • The phase information of the signal
5. How do GAN-based vocoders like HiFi-GAN improve realism?
   • By using multiple discriminators that judge the signal at different scales
   • By converting audio to text first
   • By averaging thousands of recordings together
   • By only generating low frequencies
6. In a typical modern TTS pipeline, what produces the mel-spectrogram that the vocoder consumes?
   • A noise generator
   • An acoustic model such as Tacotron 2 or FastSpeech
   • The vocoder itself
   • A microphone
7. What component in HiFi-GAN combines residual blocks with different dilation patterns?
   • A phoneme tokenizer
   • A recurrent attention decoder
   • The multi-receptive-field fusion (MRF) module
   • A speaker-verification classifier
8. What does HiFi-GAN’s multi-period discriminator help assess?
   • Whether the text transcript is grammatically correct
   • Whether the speaker’s face matches the voice
   • Whether the mel-spectrogram uses enough colors
   • Whether generated waveform segments reflect realistic periodic patterns