Bài kiểm tra có liên kết hướng dẫn · Trung bình Cấp độ

Bài kiểm tra phát hiện sự kiện âm thanh được giám sát yếu

Kiểm tra nhãn clip yếu so với ranh giới sự kiện mạnh, tổng hợp, âm thanh chồng chéo, đánh giá thời gian và chuyển đổi từ tổng hợp sang thực.

Đường dẫn hướng dẫn liên quanPhát hiện sự kiện âm thanh được giám sát yếu
Câu hỏi 1 của 8

A clip is tagged “doorbell” with no timestamp. Why is that weak supervision for detection?

Xem trước mọi câu hỏi8 câu hỏi

Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.

  1. A clip is tagged “doorbell” with no timestamp. Why is that weak supervision for detection?
    • The clip contains no waveform
    • The label gives presence but not event onset or offset
    • A doorbell cannot be recognized acoustically
    • Only one event can occur in ten seconds
  2. What additional data are needed to score a model’s proposed event timing directly?
    • Only the clip’s class tag
    • Only a model confidence score
    • Only the recording file size
    • Human-marked or otherwise trusted onset and offset references
  3. A model predicts “doorbell” correctly but at the wrong second. What does that show?
    • Clip classification and temporal localization are distinct
    • The clip label must be wrong
    • The event source was certainly isolated
    • The recording cannot contain overlapping events
  4. Why can a clip contain several positive event labels?
    • Each clip must be reduced to one class
    • The task assumes non-overlapping events
    • Sounds can overlap or occur at different moments in one recording
    • Only a single onset can be scored per recording
  5. How can a frame-level model train from a clip-level tag?
    • Assume every frame has the event label
    • Aggregate frame scores into a clip score for comparison with the tag
    • Ignore the sound waveform
    • Use the test-set event times as training labels
  6. What risk arises from training on precisely labeled synthetic mixtures?
    • The generated acoustics may not match real recordings
    • Synthetic examples have no timestamps
    • Real audio cannot contain multiple sources
    • A known mixture guarantees deployment accuracy
  7. Why specify timing tolerance or overlap rules in evaluation?
    • Every detector has a universal natural cutoff
    • A weak clip label defines the exact onset
    • Different matching rules leave event scores unchanged
    • Boundary scores depend on what counts as a match
  8. Which practice best supports a deployed detector after training with weak labels?
    • Report only clip-tagging accuracy
    • Assume every untagged sound is absent
    • Test false alarms, missed sounds and event timing on representative audio
    • Keep only the most attractive demo