YAMNet Audio Classification Quiz

Answers and explanations appear as you take the quiz.

1. What does YAMNet predict according to its documented model description?
   • Verified identities of individual speakers
   • Word-level transcripts for every recording
   • Audio event classes from the AudioSet vocabulary
   • Exact musical-note scores for every recording
2. What can YAMNet provide besides class scores?
   • Verified annotations that replace a labeled evaluation set
   • Intermediate audio embeddings
   • Speaker identities that remain verified across recordings
   • A new project-specific class ontology without any adaptation
3. Why may YAMNet fail to express a project's very specific label?
   • Every audio input must use the recording's original sample rate unchanged
   • It outputs only reconstructed waveforms rather than scores
   • An embedding makes every unsupported class directly available as a named output
   • Its output vocabulary is limited to its trained class taxonomy
4. What does a frame-level score represent?
   • A score associated with a model window, not automatically a final clip label
   • A verified label that necessarily applies to the entire recording
   • A count of every event throughout the recording
   • A guarantee that the highest class is the only audible event
5. Why is changing a sample-rate metadata field alone insufficient for resampling?
   • It automatically converts every stereo signal to mono
   • It changes the number of AudioSet output classes
   • It does not recompute the waveform at the new sampling rate
   • It removes the need to check sample values and timing
6. How can downstream teams reuse YAMNet embeddings?
   • Replace all target-task evaluation with pretraining scores
   • Train a smaller classifier for project-specific labels
   • Assume all target labels are already distinguished by the same score columns
   • Guarantee that downstream probabilities are calibrated without testing
7. What does an AudioSet score establish about a new deployment environment?
   • Guaranteed accuracy in that setting
   • A calibrated project-specific probability
   • That all event categories are represented
   • A model output that should be validated on that domain
8. Which split helps avoid leakage when recordings share a session or source?
   • Group recordings by source or session before partitioning
   • Shuffle derived crops across every split
   • Put one copy in train and another in test
   • Split only the model's output scores