Bài kiểm tra có liên kết hướng dẫn · Cứng Cấp độ

SI-SDR cho bài kiểm tra tách nguồn âm thanh

Kiểm tra tính bất biến của mức tăng SI-SDR, phép chiếu mục tiêu và phần dư, căn chỉnh tham chiếu công bằng, giới hạn điểm và kiểm tra cảm quan.

Đường dẫn hướng dẫn liên quanSố liệu tách Si Sdr
Câu hỏi 1 của 8

What simple change to an estimate does SI-SDR intentionally ignore?

Xem trước mọi câu hỏi8 câu hỏi

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

  1. What simple change to an estimate does SI-SDR intentionally ignore?
    • One overall nonzero gain change
    • Replacing a word in speech
    • A timing offset of several seconds
    • Adding a new instrument
  2. In the SI-SDR calculation, which component is compared with residual error?
    • Only the unscaled reference energy
    • Only the mixture energy before separation
    • Only listener-rated quality of the estimate
    • The estimate component projected onto the reference
  3. Why should a score report name the exact SDR variant?
    • SI-SDR and older BSS_eval SDR handle all transformations identically
    • Different definitions can treat gain or filtering differently
    • Every implementation uses the same alignment and channel rules
    • Naming the variant makes reference audio unnecessary
  4. A separator improves average SI-SDR but loses consonants. What remains to evaluate?
    • Only an even higher SI-SDR mean
    • Only the number of output files
    • Listener intelligibility and task-specific errors
    • Only the song title
  5. Why must estimate and reference be time-aligned for fair SI-SDR comparison?
    • A timing shift can create large waveform residual error
    • Timing shifts are removed by scale invariance
    • Alignment is irrelevant if two recordings sound similar
    • Only the file duration affects residual error
  6. Which report helps expose rare separation disasters?
    • Only one combined mean
    • Only training-set loss
    • Only checkpoint size
    • Per-source score distributions and listening cases
  7. Why might a quiet but shape-correct vocal estimate still score well under SI-SDR?
    • SI-SDR ignores all waveform errors
    • The allowed gain fit removes the overall level difference
    • The reference was necessarily silent
    • Quiet files cannot contain distortion
  8. Which question SI-SDR cannot answer alone?
    • Whether the model improved its own SI-SDR on a named reference
    • Whether target projection and residual are defined
    • Whether the isolated dialogue sounds natural to listeners in the intended use
    • Whether the estimate and reference are aligned