MusicLM Hierarchical Music Generation Quiz

Answers and explanations appear as you take the quiz.

1. How does MusicLM formulate conditional music generation?
   • As hierarchical sequence-to-sequence modeling over audio tokens
   • As painting pixels with a diffusion model
   • As concatenating fixed prerecorded loops
   • As simulating instrument mechanics
2. What is the role of semantic tokens in MusicLM?
   • They store the song's file metadata
   • They capture high-level structure like melody and rhythm
   • They translate lyrics into other languages
   • They control the output volume
3. Which component provides the fine acoustic detail like timbre and texture?
   • The text prompt directly
   • A Griffin-Lim reconstruction step
   • The SoundStream neural codec's acoustic tokens
   • A handwritten music theory rulebook
4. How does MusicLM connect a text prompt to musical audio?
   • By keyword matching against song titles
   • By asking a human to label each clip
   • Through a separate translation API
   • Through MuLan, a joint music-text embedding space
5. Why does the hierarchical, staged design help with long-range structure?
   • It lets coarse, slow tokens model long-term form without huge sequence lengths
   • It removes the need for any training data
   • It plays the song faster to fit more in
   • It only ever generates three seconds at a time
6. What does the MusicLM paper report about its generated music duration and sample rate?
   • Every output is exactly 30 seconds at 8 kHz
   • Its experiments report 24 kHz music that can remain consistent over several minutes
   • Only one-second clips at 48 kHz
   • Uncompressed MIDI of unlimited duration
7. What additional input can MusicLM use alongside text to guide generation?
   • A spectrogram-only image caption
   • A microphone-array steering vector
   • A hummed or whistled melody
   • A source-code repository
8. What did the MusicLM paper release as MusicCaps?
   • A collection of 5,500 synthesizer presets
   • A speech transcription benchmark
   • A set of SoundStream model weights only
   • 5,500 music-text pairs with human-written descriptions