Kvíz spojený s průvodcem · Střední úroveň

Audio Spectrogram Transformer (AST) Quiz

Test AST spectrogram patches, transformer attention, clip-level labels, preprocessing choices, timing limits, and field evaluation.

Související vodicí cestyAudio Spectrogram Transformer
Otázka 1 z 8

What representation does the original AST feed to its transformer?

Náhled každé otázky8 otázky

Odpovědi a vysvětlení se zobrazují, jak budete v kvízu.

  1. What representation does the original AST feed to its transformer?
    • Patches of an audio spectrogram
    • One summary vector of MFCC means per clip
    • A sequence of raw waveform samples without a spectrogram
    • Only hand-selected frequency peaks
  2. Why can transformer attention help on a repeating sound?
    • It guarantees the sound source is isolated
    • It removes the need for audio labels
    • It directly measures the room geometry
    • It can relate evidence from separated moments in the clip
  3. What does an AST clip-classification score not directly provide?
    • A framewise boundary for every event
    • Exact onset and offset of each sound event
    • A separated waveform for each source
    • A transcript aligned to the clip
  4. Why record sample rate and spectrogram settings when comparing AST results?
    • They prove every label is correct
    • They identify who recorded the audio
    • They change the patches and features presented to the model
    • They eliminate domain shift
  5. Benchmarkové skóre AST je vysoké, ale tovární alarm je vzácný. Co by se mělo testovat příště?
    • Per-class errors on representative factory recordings
    • Only the benchmark aggregate
    • Only checkpoint file size
    • Only the model paper’s title
  6. Why is attention weight alone not a full explanation of a tag?
    • Attention never uses any audio information
    • Every high-weight patch is an annotated event
    • It automatically gives a human transcript
    • It does not prove causal reliance on a particular audible event
  7. Which task requires additional machinery beyond an AST tagger?
    • Assigning clip-level sound categories
    • Recovering an isolated dialogue waveform from a mix
    • Scoring a held-out tagged clip
    • Fine-tuning a classifier head
  8. What deployment property should be measured on the target device?
    • Only an in-paper aggregate accuracy
    • Only training-time FLOPs on another GPU
    • Inference latency and memory alongside accuracy
    • Only the number of model parameters