NaturalSpeech and Latent Diffusion TTS Quiz

Answers and explanations appear as you take the quiz.

1. What did the original NaturalSpeech paper report for its LJSpeech evaluation?
   • Human-level quality under the paper’s stated evaluation protocol
   • Human-level quality on every language and speaker
   • Perfect equality to every human recording
   • A universal guarantee of naturalness
2. What does NaturalSpeech 2's latent diffusion model actually generate?
   • Raw audio waveforms directly
   • Continuous latent vectors from a neural audio codec
   • Printed phonetic transcripts
   • Image spectrogram pictures for human viewing
3. How does a diffusion model generate data at a high level?
   • It memorizes and replays training clips
   • It sorts samples alphabetically
   • It learns to reverse a step-by-step noising process, denoising from random noise to data
   • It predicts only the next word of text
4. What key capability does latent diffusion give NaturalSpeech 2?
   • Translating speech into images
   • Removing the need for any text input
   • Generating only a single fixed voice
   • Zero-shot voice cloning from a short reference prompt
5. What is the central idea of NaturalSpeech 3's 'factorized diffusion'?
   • Separating speech into disentangled attributes like content, prosody, and timbre, modeled independently
   • Generating speech twice as fast by skipping text
   • Compressing audio into a single integer
   • Using only rule-based signal processing
6. Why does high-fidelity diffusion TTS heighten the need for safeguards like watermarking?
   • Because it makes audio files too large to store
   • Because realistic zero-shot voice cloning raises clear misuse and impersonation risks
   • Because it only works in one language
   • Because diffusion models cannot be trained at all
7. After NaturalSpeech 2 diffusion generates codec latents, what converts them into speech?
   • A text tokenizer turns them into phonemes
   • A speaker diarization model turns them into subtitles
   • A neural audio-codec decoder turns the generated latents into waveform audio
   • A spectrogram classifier selects a genre
8. What can NaturalSpeech 3’s factorized speech representation make easier?
   • Replacing the need for any audio decoder
   • Guaranteeing every cloned voice has consent
   • Converting speech into text without an acoustic model
   • Adjusting attributes such as prosody and timbre more independently