VALL-E and Codec Language Models Quiz

Answers and explanations appear as you take the quiz.

1. What core idea distinguishes VALL-E from earlier text-to-speech systems?
   • It predicts neural codec tokens as a language modeling task instead of predicting a spectrogram
   • It uses only handcrafted signal processing
   • It requires hours of audio from the target speaker
   • It generates text from speech
2. How much reference audio does VALL-E need to clone a new speaker's voice?
   • About one hour
   • About three seconds
   • No reference audio at all
   • About thirty minutes
3. Which neural codec does VALL-E use to produce its audio tokens?
   • MP3
   • Opus
   • EnCodec
   • WAV
4. Why does VALL-E use both an autoregressive and a non-autoregressive stage?
   • To translate the speech into multiple languages
   • To separate music from speech
   • To compress the text input
   • Because codec tokens are layered, so it predicts coarse tokens sequentially and fine tokens in parallel
5. Roughly how much speech data was VALL-E trained on, enabling strong zero-shot cloning?
   • About 60,000 hours
   • About 100 hours
   • About 1,000 hours
   • About 10 hours
6. Besides timbre, what else can VALL-E preserve from the short reference clip?
   • The speaker's exact words forever
   • The acoustic environment of the recording
   • The video that accompanied it
   • The original file format
7. What does VALL-E condition on to imitate a target speaker in its demonstration setup?
   • A photograph and a written biography only
   • A long labeled transcript from every target-speaker recording
   • A short audio prompt from that speaker plus the text to speak
   • A speaker ID with no audio example
8. What text representation does the VALL-E paper use as input to its text-to-code model?
   • A raw waveform of the target utterance
   • A speaker’s face embedding
   • A list of MIDI notes
   • A phoneme sequence