Kvíz spojený s průvodcem · Těžko úroveň

Semantic Versus Acoustic Audio Tokens Quiz

Test semantic versus acoustic tokens, AudioLM hierarchy, long-range content, timbre reconstruction, error propagation, and voice privacy.

Související vodicí cestySemantic Vs Acoustic Audio Tokens
Otázka 1 z 8

In the cited AudioLM design, what are semantic tokens used to help represent?

Náhled každé otázky8 otázky

Odpovědi a vysvětlení se zobrazují, jak budete v kvízu.

  1. In the cited AudioLM design, what are semantic tokens used to help represent?
    • The precise legal source of the clip
    • An untouched waveform sample sequence
    • Longer-range content structure in audio
    • A verified word-for-word transcript
  2. What do acoustic codec tokens contribute to generation?
    • A guarantee that every word is factual
    • A legal recording consent signal
    • A direct human annotation of every note
    • Information for reconstructing timbre and fine waveform detail
  3. Why predict semantic tokens before acoustic detail in a hierarchy?
    • It eliminates every audio artifact
    • Broad structure can condition later fine-grained sound generation
    • It makes all later tokens unnecessary
    • It turns generated speech into source evidence
  4. Why is a semantic audio token not the same as a transcript word?
    • It is a learned discrete code, not a verified human word label
    • Every token is a sentence
    • It is always a raw waveform sample
    • No words can ever be inferred from audio
  5. A generated voice sounds like a speaker but says a new sentence. What conclusion is unsafe?
    • That speaker timbre may be preserved
    • That audio can be synthetic
    • That the real speaker actually said the generated words
    • That acoustic tokens can carry voice cues
  6. What can happen if an early semantic prediction is wrong?
    • Later stages are guaranteed to repair content
    • An early token cannot affect decoded audio
    • The prompt waveform changes retroactively
    • Later acoustic generation may faithfully render the wrong content
  7. Which evaluation pair covers different failure types?
    • Only file size and sample rate
    • Content coherence and perceived acoustic quality
    • Only a single transcript WER for music
    • Only model parameter count
  8. Why consider consent when continuing a person’s voice from an audio prompt?
    • Acoustic details can preserve recognizable speaker traits
    • Tokenization automatically anonymizes speech
    • Semantic codes remove all identity cues
    • A short prompt grants unlimited reuse rights