Tacotron 2 Quiz

Answers and explanations appear as you take the quiz.

1. What intermediate representation does Tacotron 2 predict from text?
   • A mel-spectrogram
   • A JPEG image
   • Raw text phonemes only
   • A MIDI file
2. What converts Tacotron 2's spectrogram into an actual waveform?
   • A spell checker
   • A modified WaveNet vocoder
   • A search engine
   • A convolutional image classifier
3. What problem does location-sensitive attention help prevent?
   • Overheating the GPU
   • Loss of internet connection
   • Repeated or skipped words during synthesis
   • Spelling mistakes in the input text
4. How does Tacotron 2 determine that it has generated the end of an utterance?
   • A human must press a button for every utterance
   • It always generates a fixed number of frames
   • The vocoder detects silence and interrupts the decoder
   • A learned stop-token prediction signals the decoder to stop
5. What overall architecture style does the text-to-spectrogram part use?
   • Sequence-to-sequence with attention
   • A simple lookup table
   • A decision tree
   • A static rule set
6. Which later model is a faster, non-autoregressive successor in spirit to Tacotron 2?
   • ResNet
   • FastSpeech 2
   • BERT
   • AlexNet
7. What does the post-net in Tacotron 2 refine?
   • The input text spelling
   • The speaker’s identity metadata
   • The predicted mel-spectrogram
   • The final transcript punctuation
8. What does FastSpeech 2 add as conditioning information for speech variation?
   • A video soundtrack transcript only
   • A random microphone serial number
   • A mandatory image embedding
   • Pitch, energy, and more accurate duration information