FastPitch Pitch-Controllable TTS Quiz

Answers and explanations appear as you take the quiz.

1. What extra signal does FastPitch explicitly predict for each input token?
   • The pitch (fundamental frequency)
   • The speaker's gender
   • The background noise level
   • The language ID
2. How does FastPitch generate the mel-spectrogram so quickly?
   • By streaming audio one sample at a time
   • Non-autoregressively, in a single parallel forward pass
   • By copying from a recorded database
   • By running on a single CPU core only
3. How can a user change the intonation contour in FastPitch output without retraining?
   • By deleting phonemes
   • By changing the random seed only
   • By scaling or editing the predicted pitch values
   • By lowering the sample rate
4. During training, how is the pitch target assigned per token?
   • A random value
   • The maximum amplitude of the audio
   • The first frame's value only
   • The average fundamental frequency over the token's duration
5. What synthesis throughput did the FastPitch paper report for a typical utterance?
   • Over 900× real time for mel-spectrogram synthesis in the reported setup
   • About one second of audio per hour
   • Exactly 2× real time for waveform output
   • No measurement of synthesis speed
6. Why is per-token pitch control described as intuitive but coarse?
   • It predicts pitch per audio sample
   • It predicts one pitch value per symbol, not per frame
   • It ignores pitch entirely
   • It only works for one speaker
7. Which earlier parallel TTS architecture is FastPitch based on?
   • Tacotron 2
   • WavLM
   • FastSpeech
   • HiFi-GAN
8. Besides pitch, what token-level signal helps FastPitch align text with output frames?
   • Speaker gender
   • Background noise level
   • Language identity only
   • Predicted token duration