Demucs Music Source Separation Quiz

Answers and explanations appear as you take the quiz.

1. What does Demucs do to a finished song?
   • Separates it into stems like vocals, drums, bass, and other
   • Compresses it into a smaller file
   • Translates the lyrics
   • Speeds it up without changing pitch
2. What makes Hybrid Demucs 'hybrid'?
   • It mixes two different songs together
   • It processes audio in both waveform and spectrogram domains
   • It uses both English and French lyrics
   • It runs half on CPU and half on paper
3. Which metric did the original Demucs paper report for separation quality on MUSDB?
   • Word error rate (WER)
   • Mean reciprocal rank (MRR)
   • Signal-to-Distortion Ratio (SDR)
   • Intersection over Union (IoU)
4. Which research organization is associated with the original Demucs project?
   • The Allen Institute for Artificial Intelligence
   • Kyutai Labs
   • The International Society for Music Information Retrieval
   • Facebook AI Research (FAIR)
5. Why did early Demucs work directly on the raw waveform?
   • To avoid losing phase information
   • Because spectrograms are illegal
   • To make files larger on purpose
   • Because waveforms contain no sound
6. Which dataset is central to training and benchmarking Demucs?
   • ImageNet
   • MUSDB18
   • MNIST
   • Common Crawl
7. What attention pattern does the HT-Demucs paper add between its waveform and spectrogram branches?
   • Attention only over lyric tokens
   • A single attention layer over stereo channels only
   • Self-attention within each domain and cross-attention across domains
   • No attention; only a nearest-neighbor lookup
8. What did the HT-Demucs paper report about extra training songs?
   • Extra songs reduced SDR by 10 dB in every setting
   • The model reached the same result with zero training songs
   • The paper reported only a word-error-rate change
   • With 800 extra songs, HT-Demucs improved SDR by 0.45 dB over Hybrid Demucs trained on the same data