StyleTTS 2 Style Diffusion Quiz

Answers and explanations appear as you take the quiz.

1. What does StyleTTS 2 model using a diffusion process?
   • A latent 'style vector' controlling prosody and timbre
   • The raw audio waveform sample by sample
   • The text transcript of the speech
   • The listener's hearing response
2. Which pre-trained speech model is used as an adversarial discriminator in StyleTTS 2?
   • ResNet-50
   • WavLM
   • BERT
   • CLIP
3. Why can StyleTTS 2 say the same sentence in many natural ways?
   • It randomly corrupts the input text
   • It uses a different language model each time
   • Style is sampled from a distribution rather than fixed
   • It changes the phoneme spelling
4. What did the StyleTTS 2 authors report on the LJSpeech listening evaluation?
   • It outperformed humans on every language and dataset
   • No listeners evaluated the samples
   • It was tested only on music
   • Native English listeners rated its samples above the human recordings in that study
5. What does 'end-to-end' training give StyleTTS 2?
   • Gradients flow from waveform quality back through the whole pipeline
   • It needs no text input at all
   • It only trains the vocoder separately
   • It removes the need for any neural network
6. What can sampling different style vectors do for the same text?
   • Change the transcript into a different language automatically
   • Produce different prosodic realizations of the written content
   • Guarantee every rendition sounds human
   • Remove the need for speech synthesis
7. What speech input does ordinary StyleTTS 2 synthesis require to sample a style?
   • A three-second reference clip is always required
   • A human listener must speak the sentence first
   • No reference speech is required; it can generate style from the text-conditioned model
   • A transcript of every training speaker
8. On which dataset did the StyleTTS 2 paper report results for zero-shot speaker adaptation?
   • ImageNet
   • COCO
   • MUSDB18
   • LibriTTS