引導式測驗 · 中等 等級

CLAP:對比式語言-音訊預訓練測驗

測試 CLAP 配對編碼器、對比對齊、文字音訊檢索、提示靈敏度、相似性限制和時間範圍。

問題 1 的 8

How does CLAP connect a written sound description with a recording?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. How does CLAP connect a written sound description with a recording?
    • One speech-to-text model that emits captions
    • A fixed audio-only classifier with no text input
    • It trains audio and text encoders into a shared embedding space
    • An isolated-source decoder conditioned on a phrase
  2. What does contrastive training encourage for a matched audio-text pair?
    • Higher similarity than mismatched pairs under the training objective
    • Guaranteed factual truth for every caption
    • Exact onset and offset of each sound
    • The same waveform as the text embedding
  3. Why can two reasonable prompts rank differently for one clip?
    • An audio clip must have only one label
    • Similarity scores are legal identity proofs
    • Prompts cannot include natural language
    • Wording and specificity affect the text embedding and comparison
  4. A clip has a high CLAP score for “dog barking.” What remains unproven?
    • That a matched event occurred at an exact second
    • That a dog bark definitely occurred at a specific time
    • That a source-separation stem can be reconstructed
    • That the written prompt was a verified human annotation
  5. Which use fits CLAP without claiming more than the model gives?
    • Certify the legal source of every sound
    • Recover isolated vocals from a soundtrack
    • Shortlist likely audio clips for a human to review
    • Diagnose a hearing condition from one score
  6. Why test rare local sounds separately?
    • Training captions may underrepresent them or use different descriptions
    • Rare sounds cannot be recorded
    • A shared embedding guarantees equal coverage
    • All languages name sounds identically
  7. What data are needed to evaluate exact sound-event timing?
    • Only a clip-level similarity score
    • Only an embedding-vector distance
    • Only a list of candidate text prompts
    • Reference onset and offset annotations
  8. What should a user retain when a CLAP match supports a public claim?
    • Only a similarity number
    • The original recording and a reviewable comparison
    • Only the query text
    • Only the model family name