FastSpeech and Non-Autoregressive TTS Quiz

Answers and explanations appear as you take the quiz.

1. What does 'non-autoregressive' mean in the context of FastSpeech?
   • It generates all spectrogram frames in parallel rather than sequentially
   • It generates audio frames one at a time in sequence
   • It cannot handle long sentences
   • It requires no training data
2. Which problem of autoregressive models like Tacotron 2 does FastSpeech specifically address?
   • Inability to produce any audio
   • Slow generation and skipped or repeated words from attention errors
   • Lack of support for English text
   • Requiring quantum hardware
3. What is the role of the 'length regulator' in FastSpeech?
   • It compresses the final audio file
   • It removes silence from recordings
   • It repeats each phoneme's representation according to its predicted duration to match audio length
   • It translates phonemes into text
4. Which additional speech variations does FastSpeech 2 condition on during training and inference?
   • Only speaker age and recording location
   • Video frames and subtitle color
   • A single fixed global loudness value
   • Pitch, energy, and duration information
5. Why does FastSpeech's inference time barely grow with sentence length?
   • All frames are generated in parallel rather than sequentially
   • It ignores most of the input text
   • It only ever outputs one-second clips
   • It uses a lookup table of pre-recorded sentences
6. How does the length regulator use predicted durations in FastSpeech?
   • It selects the final audio codec
   • It expands phoneme-level representations to align with the target spectrogram length
   • It estimates the speaker’s identity from video
   • It converts text into a sequence of labels for translation
7. What is one tradeoff of non-autoregressive spectrogram generation?
   • It guarantees identical prosody to an autoregressive model
   • It removes the need for a waveform vocoder
   • Parallel output can be faster, but duration and speech variation must be modeled explicitly
   • It cannot accept text as input
8. What training target does FastSpeech 2 use instead of FastSpeech’s teacher-generated simplified outputs?
   • Only predicted durations with no acoustic target
   • A translated text transcript
   • The output of a GAN discriminator only
   • Ground-truth target speech representations