Викторина о музыкальном разделении

Ответы и пояснения появляются по мере прохождения теста.

1. What does a source-separation system produce from a mixed recording?
   • Estimates of component tracks or stems
   • Perfect copies of the original studio tracks
   • A new composition with no overlap
   • A legal license for remixing
2. Why may a stem sound worse when isolated than in the final mix?
   • The original recording changes rights
   • Leakage and artifacts can be more obvious after isolation
   • The model always removes the bass
   • The export format restores missing detail
3. A model is trained for a few broad stems. What should a user not assume?
   • That overlap can be ambiguous
   • That it returns broad categories
   • That it can isolate every individual instrument
   • That settings affect results
4. How should two separation models be compared fairly?
   • Amplify one output more than the other
   • Use different songs and settings
   • Compare only the model names
   • Use the same source, export format, and processing settings
5. What do objective metrics need for meaningful comparison?
   • A commercial license
   • Reference stems
   • A video camera
   • A model without a checkpoint
6. Why record the exact checkpoint used?
   • It removes leakage
   • A checkpoint grants publishing rights
   • Different versions may behave differently
   • It changes the original mix
7. What should be checked before publishing a remix made from separated audio?
   • Permissions for remixing, redistribution, or publication
   • Only whether the vocal sounds clean
   • The model’s inference speed
   • Whether the output is a WAV file
8. A vocal stem is made for a language-learning exercise. What should the editor assess?
   • Whether a listener can identify the artist
   • Only whether the file contains vocals
   • Количество параметров модели
   • Missing consonants, residual instruments, and artifacts against the original