引導式測驗 · 中等 等級
MUSDB18 音樂分離基準測驗
測試 MUSDB18 四干參考、100/50 訓練測試分割、標準與 HQ 變體、每幹報告和傳輸限制。
問題 1 的 8
Which four reference stems are named in MUSDB18?
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。
- Which four reference stems are named in MUSDB18?
- Dialogue, music, effects and silence
- Vocals, drums, bass and other
- Left, right, center and surround
- Pitch, tempo, rhythm and loudness
- How many songs are in MUSDB18’s standard training and test split?
- 50 training and 100 test tracks
- 150 training and none held out
- One track per stem category
- 100 training and 50 test tracks
- Why report errors separately for vocals and drums?
- One stem can improve while another gets worse
- All stem errors are identical
- A mean guarantees no source leakage
- Only vocals have reference audio
- What should a paper state when using MUSDB18-HQ?
- That HQ is the identical encoded release
- That test songs became training songs automatically
- That it used the distinct high-quality WAV variant
- That an HQ score certifies live recordings
- A model is tuned repeatedly on the public test songs. What risk follows?
- The test set becomes larger automatically
- The final score may reflect overfitting to that test set
- The stem labels change
- The model can no longer separate audio
- Why is the “other” stem difficult to interpret as one source?
- It combines a range of remaining instruments and sounds
- It contains only one fixed instrument
- It is always pure silence
- It has no waveform reference
- What extra test helps before launching a live-concert separator?
- Only the MUSDB18 training loss
- Only one studio-song demo
- Only the model architecture diagram
- Evaluation on representative live recordings and listener review
- What does access to a MUSDB18 mixture not grant automatically?
- A reference for evaluation
- A way to compare model output with stems
- Rights to publish separated derivatives of every underlying recording
- A shared research benchmark