Maskowanie czasowo-częstotliwościowe w quizie dotyczącym separacji źródeł

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. How does a time-frequency mask produce a target estimate from a mixture STFT?
   • Use the clean target spectrogram as the output without consulting the mixture
   • Apply one constant gain to the entire recording
   • Multiplies the mixture representation by bin-wise weights
   • Assign a single source label to the whole clip
2. What values characterize an ideal binary mask in a simple two-source assignment?
   • A continuously varying real weight across the unit interval
   • Zero or one
   • A complex-valued correction with arbitrary phase
   • The same gain for every time-frequency bin
3. How can a ratio mask differ from a hard binary mask?
   • It must assign every bin exactly zero or one
   • It guarantees that no interference remains
   • It ignores relative source strength when defining its weights
   • It can represent graded contributions with soft values
4. What can a complex mask adjust that a real magnitude mask usually does not?
   • Phase as well as magnitude
   • Only magnitude, never phase
   • Only the mixture phase, with no magnitude change
   • Only a scalar gain applied to the whole waveform
5. Which phase is often reused when a real-valued magnitude mask reconstructs an overlapping target?
   • The clean target's original phase supplied independently
   • A separately estimated target phase for every real-valued mask
   • The mixture phase
   • Zero phase imposed on every bin
6. Why should a ratio-mask definition be stated explicitly?
   • All valid ratio masks are hard binary decisions
   • Different methods define ratios using magnitudes or powers
   • Ratio masks are computed from logarithms only
   • Every ratio mask uses exactly the same exponent and target definition
7. What should match between masking and waveform reconstruction?
   • Any hop size works if the final duration is forced to match
   • The synthesis window is irrelevant after masking
   • The inverse transform can infer every analysis convention from mask values alone
   • Compatible STFT window and hop settings
8. Which evaluation approach is useful for separated speech?
   • Combine appropriate signal metrics with listening or downstream tests
   • Treat a visually smooth mask as proof of clean separation
   • Report only the training objective on fitted mixtures
   • Assume waveform additivity guarantees perceptual quality